Professional Documents
Culture Documents
Doctoral Thesis
Author(s):
Psaltopoulos, Georgios
Publication Date:
2011
Permanent Link:
https://doi.org/10.3929/ethz-a-006383560
Rights / License:
In Copyright - Non-Commercial Use Permitted
This page was generated automatically upon download from the ETH Zurich Research Collection. For more
information please consult the Terms of use.
ETH Library
Diss. ETH No. 19550
A dissertation submitted to
ETH ZURICH
presented by
GEORGIOS PSALTOPOULOS
M.Sc. TU Munich,
Dipl.-Ing., Aristotle University of Thessaloniki
born April 7, 1980
citizen of Greece
2011
Day of Doctoral Examination: February 22, 2011
Κι ἂν πτωχικὴ τὴν βρεῖς, ἡ ᾿Ιθάκη δὲν σὲ γέλασε.
῎Ετσι σοφὸς ποὺ ἔγινες, μὲ τόση πεῖρα,
ἤδη θὰ τὸ κατάλαβες οἱ ᾿Ιθάκες τὶ σημαίνουν.
And if you find her poor, Ithaca won’t have fooled you.
Wise as you will have become, so full of experience,
you will have understood by then what these Ithacas mean.
Ithaca by C. P. Cavafy
ii
Abstract
In this thesis we introduce a new paradigm for low-cost/low power wireless systems:
nonlinear MIMO systems. This new type of MIMO systems offers a way to reduce
both cost and power consumption of legacy MIMO systems by using nonlinear de-
tection methods, i.e., amplitude and phase detection. Conventional MIMO systems
studied so far strictly employ linear receivers, meaning that multiple ideal I/Q receiver
chains process the signal on every receive antenna. This renders MIMO receivers expen-
sive and power-hungry. On the other side, amplitude and phase detection are known
as low-cost approaches to receiver design, since they can be implemented with much
simpler and more energy-efficient circuitry. In fact, amplitude detection can be simply
implemented using an envelope detector, without the need for a precise oscillator ref-
erence and downconversion. Phase detection on the other hand can be implemented
with a cheap limiting amplifier and a phase/frequency discriminator. Both receivers
avoid the extravagant I/Q architecture. We investigate the combination of nonlinear
receivers with MIMO systems. Nonlinear MIMO systems merge the high-performance
world of MIMO systems with the low cost/low power world of nonlinear receivers.
Nonlinear MIMO systems have not been considered so far. The nonlinear nature of
such systems introduces certain idiosyncrasies, apart from the aforementioned advan-
tages. Linear MIMO systems have been extensively studied and are well understood,
amongst other, because their linearity property enables easy mathematical tracking.
However, mathematical tractability in nonlinear MIMO systems is not straight-forward
anymore. At the same time, most known results/observations for linear MIMO systems
are no more valid. Hence, nonlinear MIMO systems offer a new unexplored field for
research, which at the same time is very relevant to future needs. Nonlinear MIMO
systems are especially lucrative for applications that require high numbers of very low-
cost and low-power devices. The increasingly popular wireless sensor networks are a
typical example: very cheap nodes with long lifetime are crucial for the purpose of
the network. In general, in the upcoming ubiquitous wireless world, the need for low
cost/low power solutions is inexorable.
iii
Abstract
iv
Kurzfassung
v
Kurzfassung
langer Lebensdauer sind äusserst wichtig für die Zwecke dieser Netzwerke. In einer
zukünftigen, allgegenwärtigen, drahtlosen Welt sind solche kostengünstigen und leis-
tungseffizienten Lösungen unvermeidlich.
In dieser ersten Arbeit über nichtlineare MIMO Systeme erkunden wir die funda-
mentalen Eigenschaften und bestimmen das Potenzial solcher Systeme. Wir begin-
nen mit theoretischen Untersuchungen über nichtlineare MIMO Systeme, erst ausge-
hend von informationstheoretischen Aspekten und dann in die Richtung der Imple-
mentierung solcher Systeme. Zuerst zeigen wir, dass die nichtlinearen MIMO Sys-
teme, gleich wie die linearen MIMO Systeme, räumliche Freiheitsgrade ausnützen, je-
doch auf beschränkte Weise. Dies führt dazu, dass die Kombination mehrerer An-
tennen mit nichtlinearen Empfängern die Leistung mit minimalen zusätzlichen Kosten
und Energieverbrauch erhöht. Anschliessend untersuchen wir den ML-Detektor für
nichtlineare MIMO Systeme im Zusammenhang mit der Empfangsdiversität und den
räumlichen Modulationsalphabeten. Es stellt sich heraus, dass sich die nichtlinearen
MIMO Systeme anders als die linearen MIMO Systeme verhalten. Weiter entwickeln
wir mehrere Kanalschätzer, die kohärente Detektion in nichtlinearen MIMO Systemen
ermöglichen. Dies vervollständigt die nötigen Basisblöcke für ein funktionstüchtiges
System. Schliesslich verifizieren wir die Gültigkeit und das Potential der nichtlinearen
MIMO Systeme durch das Experiment an einen Referenzdesign eines MIMO Amplitu-
dendetektors.
vi
Acknowledgments
I would like to thank Prof. Dr. A. Wittneben, for guiding me throughout this thesis.
He kept inspiring and motivating me for working on nonlinear MIMO, and he provided
excellent conditions at the Wireless Communications Group as well. I would also
like to thank Prof. Dr. Ch. Mecklenbräuker for being the co-referee for this thesis.
However, I would not have started a PhD at first place if it were not for my Diploma
and Master thesis supervisors, who nurtured my appetite for doing research. Hence, I
would also like to thank Dr. M. Joham, Dr. Ch. Dimakis and Dr. K. Koutsouvelis for
their invaluable guidance during these first steps.
I would like to especially thank my office mates during the last five years, Celal
Esli and Christoph Steiner, for the great time we had during our postgraduate studies.
Besides sharing an office, we had numerous scientific discussions, we supported each
other and helped through all highs and lows that accompany a PhD. I am also grateful
to all other members of the Wireless Communications Group with whom we had lots
of fun in and outside the office.
Special thanks go to the Greek student community in Zurich for all the great time
we got to spend together. I was lucky to establish some very good friendships during
this time.
Finally, I would like to thank my family for their love and support. Recognizing they
are the most important persons in my life, I would not have completed this thesis if it
was not for them!
vii
Acknowledgments
viii
Contents
Abstract iii
Kurzfassung v
Acknowledgments vii
1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Advances in MIMO Technology . . . . . . . . . . . . . . . . . . 2
1.1.2 Low-cost Low-power Systems . . . . . . . . . . . . . . . . . . . 3
1.2 Nonlinear MIMO Systems . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Contributions and Thesis Outline . . . . . . . . . . . . . . . . . . . . . 5
I From Theory... 9
2 System Model 11
2.1 Linear MIMO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2 Nonlinear MIMO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
ix
Contents
5 Channel Estimation 59
5.1 Channel Estimation for Amplitude Detection . . . . . . . . . . . . . . . 59
5.1.1 Sufficient Channel State Information . . . . . . . . . . . . . . . 60
5.1.2 Scalar Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.1.2.1 Maximum Likelihood Estimation . . . . . . . . . . . . 63
5.1.2.2 Bayesian Estimation . . . . . . . . . . . . . . . . . . . 64
5.1.2.3 Method of Moments Estimation . . . . . . . . . . . . . 65
5.1.3 MIMO Amplitude Detection Channel Estimation . . . . . . . . 67
5.1.3.1 Simple Exhaustive Estimation Scheme (SEES) . . . . . 68
5.1.3.2 Geometrically Assisted Estimation Scheme (GAES) . . 69
5.1.3.3 Component GAES algorithm . . . . . . . . . . . . . . 72
5.1.3.4 Antenna Reference Estimation Scheme (ARES) . . . . 75
x
Contents
xi
Contents
Notation 173
Acronyms 175
Bibliography 179
Index 185
xii
Chapter 1
Introduction
1.1 Motivation
A universal goal of all communications systems is to offer the best possible perfor-
mance and quality of service while keeping cost, complexity and power consumption
at a minimum. However, satisfying the above constraints concurrently is impossible,
since the laws of physics impose a fundamental tradeoff between performance and
cost/complexity. The various levels of performance, complexity, power efficiency and
cost create an ample design space that is governed by the above tradeoff. The engi-
neers designing a specific system have the responsibility of finding a suitable balance
according to the application considered. During the design of a wireless system, the
most important/crucial characteristic is identified. The communications system is then
designed such that this characteristic is guaranteed at a certain level, and the other
design parameters are optimized according to the remaining degrees of freedom. Hence,
the most sophisticated technology would be chosen for a high-performance link that
has access to the power supply network. In this case, reducing the power consumption
would have secondary priority. Similarly, the power efficiency would be of absolute
priority in deep space communications, where access to an energy source is severely
limited.
In this context, we constrain ourselves to a specific part of the above design space
and identify possibilities to influence the fundamental tradeoff that is currently dic-
tated by the state of the art. We are interested in low cost/low power systems that
are characterized by loose performance requirements. Such systems are increasingly
important and relevant in a world where wirelessly connected devices become ubiq-
uitous. The most prominent example is sensor networks, composed of thousands of
1
Chapter 1 Introduction
simple and cheap nodes that rely on an external power supply. In this thesis we pro-
pose a new family of systems called nonlinear MIMO systems. These enable the design
of more powerful low cost/low power wireless systems, by leveraging the high perfor-
mance gains of Multiple Input Multiple Output (MIMO) systems and the simple and
power-efficient design characteristics of nonlinear receivers. Before introducing nonlin-
ear MIMO systems, we revise the advances in MIMO technology, and the state of the
art in low power/low complexity systems.
2
1.1 Motivation
to the multiple received streams that have to be processed and due to the high transfer
rates involved. This shortcoming has spurred the development of algorithms that try to
reduce the complexity. Although advances in chip design suggest that complex schemes
will be less of a problem in the future, high complexity increases the cost in any case.
The Achilles heel of MIMO systems is however the hardware requirements. First of
all, a separate receiver chain is necessary for every receive antenna. Next, the theo-
retical analysis of MIMO systems and all developed schemes assume a perfect linear
system model [3]. This implies a high quality linear In-Phase/Quadrature Phase (I/Q)
receiver chain on every receiver branch. Hence, the analog front-end includes multi-
ple low-noise amplifiers, filters, mixers and analog-to-digital converters [7]. Solely the
Local Oscillator (LO) used by the mixer in downconversion can be shared between all
receiver branches. Furthermore, it is expected that all hardware components possess
good linearity properties and high dynamic range, that the I/Q imbalance is very small
and the phase noise is controlled through a Phase-Locked Loop (PLL). These require-
ments lead to very costly and power-hungry analog front-ends, but are necessary in
order to achieve the high rates expected in theory. Although technological improve-
ments in integrated circuits will eventually reduce cost and power consumption, the
fundamental need for several I/Q branches can not be circumvented. Concluding we
can say that fundamental flaws of MIMO systems are the increased power consump-
tion and the high cost and complexity. For these reasons, MIMO technology has been
restricted to high-performance costly systems, while deployment in low cost/low power
systems has been hindered.
As discussed above, MIMO systems are not ideal for low cost/low power applications.
We explore the characteristics of low-cost/low-power systems using a typical example,
namely sensor networks. Wireless sensor networks have been a hot research topic for
over a decade. Such networks, composed of simple sensing nodes, densely deployed, can
have a multitude of applications, spreading from health, home to military. Research on
sensor networks has been focused on all communication levels, from routing protocols
down to the physical layer. Design and optimization of sensor nodes is always connected
to stringent power and complexity constraints which must be fulfilled. Since the number
of nodes can be, and in many applications must be, very high, the extreme low cost of
these devices is of paramount importance. Low cost implies limited complexity in terms
of processing power and communications technology. On the other side, the nodes are
3
Chapter 1 Introduction
mostly powered by batteries, in which case the power-efficiency defines the lifetime of
the sensor. Hence, extremely low power consumption is also a design target [8].
In order to assure low power consumption and modest complexity, nonlinear mod-
ulation techniques are very popular in realistic sensor node design. Modulations like
On-Off Keying (OOK) [9] and Frequency Shift Keying (FSK) [10, 11] both require an
envelope detector which detects the presence of the signal by measuring the received
energy. Both techniques lead to significant power savings in transmitter as well as
receiver design by avoiding power-demanding components, like mixers, LOs, PLLs,
etc., which are required in an I/Q architecture [11]. Besides designs that are based
on envelope detectors, another low-power solution is using constant envelope modula-
tions, like Constant Phase Modulation (CPM) [12]. A receiver for a CPM signal can
be implemented using a limiting amplifier, since the amplitude of the signal carries no
information. Furthermore, demodulation can be performed using a phase/frequency
discriminator. A receiver that uses a limiting amplifier and a phase/frequency dis-
criminator is orders of magnitude cheaper and requires significantly less power than a
linear I/Q receiver [13]. The common denominator for envelope detectors and phase
discriminators is non-coherent detection. In this case, non-coherent implies the absence
of a LO and a PLL that track the carrier frequency and phase. Another important
aspect for power savings in sensor networks is the short start-up time and the associ-
ated power during this step (cf. [11]). This is very important, since wireless sensors
usually operate in a low-duty cycle mode, enabling power savings while adapting to
their sporadic communication needs. Nonlinear modulation and the required receivers
are much more suitable for achieving short start-up times and saving power when the
circuitry is in stand-by.
Obviously, MIMO systems that employ several I/Q receiver branches are not suitable
for such applications. In fact, the power consumption of typical MIMO systems is
orders of magnitude higher than that of efficient sensor nodes [7]. Nevertheless, the
increased MIMO rates would be useful in sensor nodes, which could complete their
emissions faster and then go back to stand by. In terms of using MIMO techniques to
increase power efficiency, various solutions have been proposed in literature that form
virtual MIMO systems through node cooperation [7,14,15]. The central aim of all these
schemes is taking advantage of the high MIMO rates and thus reducing the transmit
duration, using e.g. distributed space time codes. However these schemes require node
cooperation in order to disseminate information locally, and synchronization in order to
transmit/receive information jointly. Furthermore, continuous rate adaptation is used
4
1.2 Nonlinear MIMO Systems
5
Chapter 1 Introduction
nonlinear MIMO systems are not known. The nonlinear nature of such systems intro-
duces certain idiosyncrasies, apart from the aforementioned advantages. Mathematical
tractability is not straight-forward anymore, while most known results/observations
for linear MIMO systems are not valid any more. Hence, nonlinear MIMO systems
offer a new unexplored field for research, which at the same time is very relevant to
future needs for low cost/low power solutions. Throughout the thesis we will consider
point-to-point nonlinear MIMO links, taking only physical layer aspects into account.
We will investigate the two most popular nonlinear receivers: amplitude-only 1 detec-
tors and phase-only detectors. Furthermore, we assume that the receiver has either
perfect channel knowledge, or has access to a channel estimate. This is often described
as a coherent reception in recent literature [3], however, it should not be confused with
coherent receivers that require a LO and a PLL. In the course of the thesis we will
answer the following questions:
1. How do nonlinear MIMO systems leverage the spatial degrees of freedom? Can
they achieve spatial multiplexing gains, similar to linear MIMO systems? Do
nonlinear MIMO systems achieve significantly higher rates compared to SISO
nonlinear systems?
2. What diversity order can the Maximum Likelihood (ML) detector for nonlinear
MIMO systems achieve? What modulation alphabets are suitable? Is the diver-
sity order affected by the modulation alphabet?
3. How can nonlinear MIMO systems estimate the Channel State Information (CSI)
required by the ML detector? Can such systems estimate the required CSI relying
only on the nonlinear receivers?
The thesis is composed of two parts. Most chapters are included in Part I, where we
develop the theory for nonlinear MIMO systems. This includes information theoretic
aspects, ML detection and spatial multiplexing, and finally channel estimation schemes
for nonlinear MIMO systems. In this part we cover both MIMO amplitude and phase
detection in parallel. In Part II we present the MIMO ampLitude dEtection (sMILE)
test bed, which is the first test bed that demonstrates detection of spatially multiplexed
streams using amplitude detectors. The results of our work have been published in
[19–25]. The content of each chapter is summarized below:
1
Throughout the thesis, we will use the terms ‘amplitude’ and ‘envelope’ interchangeably.
6
1.3 Contributions and Thesis Outline
Chapter 3: Our first investigation of nonlinear MIMO systems starts with the infor-
mation theoretical limits and characteristics of such systems. We numerically compute
achievable rates and characterize the spatial multiplexing gains of nonlinear MIMO,
compared to linear MIMO systems. As a first step we assume the receiver has perfect
CSI. We show that nonlinear MIMO systems benefit from the spatial dimensions of the
medium and achieve at least half the spatial multiplexing gain of linear N N MIMO
systems. It turns out that phase detection achieves higher rates compared to amplitude
detection. Surprisingly, nonlinear MIMO systems can take advantage of more spatial
degrees of freedom when the number of receive antennas increases. Finally, we inves-
tigate the robustness of the achievable rates to noisy CSI, and show that the behavior
of nonlinear MIMO systems is very similar to linear MIMO.
7
Chapter 1 Introduction
possible with a nonlinear receiver, but it is also not required. In the case of amplitude
detection, we introduce the notion of sufficient CSI, and devise three algorithms that
can estimate it. The three novel estimation schemes, SEES, GAES and ARES are
developed on fundamentally different bases. The performance and scaling of the pilot
length with the number of transmit antennas are different between the three schemes.
We provide guidelines regarding which scheme should be employed under certain cir-
cumstances. In Chapter 6 we will implement the three schemes in a real system. On
the other hand, channel estimation using phase detection proves to be more challeng-
ing, since complete knowledge of the complex-valued channel coefficients is required in
this case. We propose three estimation schemes for the MIMO phase detector, DES,
SEES and ARES.
Chapter 6: This chapter constitutes the second part of the thesis and deals with the
implementation of the sMILE test bed. In this chapter we describe the design of the
sMILE nodes and characterize them through measurements. A model of the sMILE
hardware is devised which will assist the analysis and signal processing. Next, the
software receiver is described, which is a software implementation of the digital signal
processing on the sMILE receiver. The basic operations of the receiver, synchroniza-
tion, channel estimation and ML detection are described. Finally we perform several
measurements and discuss the performance characteristics of the test bed. Throughout
this chapter, connections are drawn between the theoretical analysis in the first part
of the thesis and the practical implementation of such a system.
Chapter 7: At the end of the thesis we summarize the results from all Chapters
and discuss the findings and contributions. We finally highlight some topics that we
consider interesting open questions for further research.
8
Part I
From Theory...
9
Chapter 2
System Model
In Part I of the thesis we consider a system model that captures the nonlinearity in an
idealized way. The generic MIMO system is depicted in Fig. 2.1, with NT transmit and
NR receive antennas. We consider a discrete time baseband model and assume that the
sampled signal at the receiver constitutes a sufficient statistic. The transmitter emits
the signal s P CNT over the stationary memoryless flat fading channel H P CNR NT ,
with tap-gain Hij from the j-th transmit to the i-th receive antenna. We use a block
fading model for the channel H. The channel remains constant for some (coherence)
period, long enough to allow for accurate estimation, and changes to an independent
realization in the next block. This can be achieved with a sufficiently long interleaver.
The received vector z P CNR is perturbed by a zero-mean circularly symmetric Gaussian
noise vector w P CNR , with autocorrelation function E rwk wlH s σw2 INR δk,l , where k
and l are symbol instants. A function g pq acts on every antenna, producing the output
vector y g pr q g pz w q. The detector has access only to y.
g pq
g pq
g pq
s H z r y
w
Figure 2.1: Generic Nonlinear MIMO System Reference Model
11
Chapter 2 System Model
In legacy linear MIMO systems, the following simple linear input-output relationship
is used:
y r Hs w, (2.1)
i.e., the function g pq does not modify the received signal. The receiver has access to the
complex-baseband signal that is obtained by I/Q receivers on every receive antenna.
The linearity of (2.1) enables easy mathematical manipulation and analysis. However,
this simple system model neglects several sources of imperfection which exist in real
communication systems, such as I/Q imbalance, amplifier nonlinearities, phase noise
and other Radio Frequency (RF) imperfections [26]. In fact, performance of real MIMO
systems is severely affected by these imperfections and falls short of the theoretically
expected performance.
For nonlinear MIMO, the function g : CNR ÞÑ RNR maps the complex-valued received
signal r from the complex plane into one real dimension, element-wise on every antenna.
This nonlinear processing allows for a very simplified receiver structure, which consumes
significantly less power compared to a linear MIMO I/Q receiver. In particular, we
consider two types of nonlinear functions known from literature and motivated by their
efficient implementation: an amplitude and a phase function. These are described
as follows:
a
yi,ampl. gampl.priq |ri| ℜtri u2 ℑtri u2 P R , (2.2)
yi,phase gphase pri q =pri q atan2 pri q P pπ, π s, (2.3)
12
2.2 Nonlinear MIMO
Throughout the thesis, we will use argpq or =pq for indicating the principle argument
of a complex number, given by the atan2pq function. For both nonlinear functions, the
main characteristic is that the signal y is missing half the dimensions after processing
with the function g. Throughout the following chapters, the probability of the output
y conditioned on z will play a crucial role. Since the noise vector w is white across
the receive antennas and the nonlinear function g pq acts on each antenna separately,
the conditional distribution of y given z can be factorized as
¹
NR
f py |z q f pyi |zi q, (2.5)
i 1
where yi and zi are the ith elements of the vectors y and z, respectively. When z is
given, r is complex Gaussian distributed like CN pz, σw2 INR q. The amplitude yi |ri |
is Ricean distributed [27]
where ρi |zi|2{σw2 and ∆φi rpyi,phase =pzi qq mod 2πs P pπ, πs1. I0pq is the zeroth
order modified Bessel function of the first kind and erfcpq is the complementary error
function.
1
The mod 2π operator is understood to mean ‘over some 2π interval’. In our case, we choose the
interval pπ, π s for convenience (cf. [29]).
13
Chapter 2 System Model
1.8
2
1.8 σw = 0.1
2
σ2 = 10 1.6 σ =1
w
w
1.6
2 σ2 = 10
σw = 1 w
1.4 2
σw = inf
1.4 2
σw = 0.1
1q
1q
1.2
1.2
0.8
0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
0 1 2 3 4 5 6 7 8 9 10 −3 −2 −1 0 1 2 3
yi ∆φi
(a) (b)
Figure 2.2: (a) Probability density function of the amplitude yi , (b) Probability density
function of the phase ∆φi .
The two pdfs are depicted in Figs. 2.2a,2.2b for amplitude and phase, respectively, for
zi 1 and various values for σw2 . For σw2 Ñ 0 both distributions converge to Gaussian
distributions. For σw2 Ñ 8 the Rician distribution converges to a Rayleigh distribution,
and the phase distribution converges to a uniform distribution in pπ, π s. Note that
although at low σw2 the distributions converge to Gaussians, the support of the pdfs is
not p8, 8q.
14
Chapter 3
For this chapter only, we distinguishing between Random Variables (r.v.’s) and real-
izations of r.v.’s using different notation. Thus, bold-faced italic lower and upper case
letters stand for random vectors and matrices, respectively, while their realizations
are denoted with upright letters. bi is the ith element of the vector b, while bi is a
realization of the random vector b, indexed with i.
First, we consider the case when the receiver has perfect knowledge of the channel
realization. Although not realistic, this assumption is important as a first step for
understanding nonlinear MIMO, and in order to compare with linear MIMO systems,
where the case of perfect receive CSI is well understood.
15
Chapter 3 Information Theoretic Limits
The capacity of linear MIMO systems with a total transmit power constraint P and
perfect receive CSI has been characterized in [1, 2]. For a stationary memoryless chan-
nel, the ergodic capacity is given by
Clin EH log det INR
SNR
NT
HH H , (3.1)
and the capacity achieving input distribution for i.i.d. Rayleigh fading is circularly
symmetric complex Gaussian, s CN p0, σs2 INT q. The average SNR per receive an-
tenna is defined as
2
SNR NσT2σs ,
P
(3.2)
σw2 w
with σh2 1. In the high SNR regime, the capacity is approximately given by
N¸
min
Clin Nmin log SNR
NT
E log λ2i , (3.3)
i 1
where Nmin minpNT , NR q and λi ’s are the singular values of H. Linear MIMO
systems exploit Nmin degrees of freedom, or else, achieve Nmin spatial multiplexing
gain, meaning that capacity increases by Nmin bits for every 3 dB SNR increase in the
high SNR regime [3].
Since we assumed that the channel H is memoryless, the capacity is given by the
maximum mutual information between the channel input s and the detector output
y. Furthermore, since H H is known to the receiver, the channel output is the pair
py, H q (cf. [2]). The mutual information can be written as
Ips; y, H q E H rIps; y |H Hqs , (3.4)
16
3.1 Rates with Perfect Channel State Information
We used the fact that hpy |z, Hq hpy |z q. We need the distributions f py |Hq and
f py |z q. The first distribution is known only in some cases. Otherwise, we estimate it
numerically, as explained in the sequel. The second distribution is known analytically
and can be found in (2.5),(2.6) and (2.7).
In the following we discuss how to numerically compute achievable rates for fixed
input distributions.
The integrals in (3.6) are computed using Monte Carlo (MC) integration [30]. MC
methods are especially important for the computation of multi-dimensional integrals,
i.e., more than 4 dimensions, where the traditional numerical methods become com-
putationally prohibitive. Appendix A presents a short explanation of MC integration.
Due to (2.5), (2.6) and (2.7), the distribution f py |z q appearing in the second integral
of (3.6) is known analytically for both amplitude and phase detection. To compute
this integral we create N realizations of z and y pairs, pzi , yi q, i 1, . . . , N which are
distributed according to their joint distribution f pz, y q. That is, zi is drawn from a
distribution f pz q of our choice, e.g. by choosing a Gaussian input s, and yi is drawn
from the distribution f py |zi q. Then, the MC estimate of the conditional entropy hpy |z q
is given by
1 ¸
N
hpy |z q log pf pyi |zi qq . (3.7)
N i1
17
Chapter 3 Information Theoretic Limits
The distribution f py |Hq is necessary for the computation of hpy |Hq. However, f py |Hq
is generally not known. For a Gaussian input alphabet s and a channel realization H,
r Hs w is Gaussian with r |H CN p0, σs2 HHH σw2 I q. For amplitude detection,
f py |Hq f pg pr q|Hq is a multivariate Rayleigh distribution, while for phase detection it
is a multivariate distribution with uniformly distributed marginals. However, closed-
form expressions exist only for specific values of NR . For NR 1, y is Rayleigh or
uniformly distributed in pπ, π s for amplitude and phase detection, respectively. The
corresponding entropies are known. For NR 2, the amplitude and phase distributions
are given by [31], [32]
f phase py |Hq
|Φ| 1 λ cos1 λ ,
1 λ2 p1 λ2 q3{2
(3.9)
8π 2 Φ Φ 11 22
and χij =pΦij q. An expression exists for the trivariate amplitude distribution (see
[31], [33]). The amplitude distribution for NR 3 is given by [31], [33]
8̧
rεmp1qmIm p2|Φ12|y1y2qImp2|Φ23|y2y3q
m0
Im p2|Φ31 |y3y1 q cospmpχ12 χ23 χ31 qqs , (3.11)
where εm is the Neumann constant (ε0 1,εm 2 for m ¡ 1). However, it consists of
an infinite sum of alternating terms and proves to be numerically unstable. No closed-
form expression exists for the trivariate phase distribution. Hence, for NR ¥ 3, we
will use MC integration to estimate the probability density of certain realizations yi .
First, we create N realizations yi , distributed according to f py |Hq, for a given channel
realization. We do this by passing samples of s through the system for a given H.
Depending on NR , we evaluate the probability density from (3.8), (3.9), or numerically
as follows:
»
1 ¸
M
f pyi q f pz q f pyi |z qdz f pyi |zj q, (3.12)
M j 1
18
3.2 Rates with Noisy Channel State Information
where the samples zj are generated according to the distribution f pz q. Note that
z CN 0, σs2 HHH . Finally, the MC estimate hpy |Hq is given by
1 ¸
N
hpy |Hq logpf pyi qq. (3.13)
N i1
At the end, we average (3.7) and (3.13) over many channel realizations H.
Eq. (2.6) is used in (3.7), (3.12) as well as in the next section. However, when the
argument of the Bessel function in (2.6) is large, (2.6) becomes numerically unstable:
the Bessel function grows to infinity and the exponential term converges to zero. This
happens almost always when σw2 is small, that is, at high SNR. We use the following
approximation for (2.6) in these cases
pyi |σw2zi |q
2
σw2
f ampl. pyi |zi q 1
yi
π |zi |σw2 16yi |zi |2
e ... (3.14)
H H
x E (3.15)
is uncorrelated with the estimate H,x following a property of MMSE estimation. The
estimation error variance σe2 E r|Hi,j |2s E r|H
xi,j |2s i, j, is a measure of the quality
of the estimation. The entries of Hx are also i.i.d. zero-mean circularly symmetric
complex Gaussian with variance σh2 σe2 .
We will consider two different scenarios in our simulation results. In one case, σe2
is constant for all SNR. This corresponds to a scenario where the channel estimate
19
Chapter 3 Information Theoretic Limits
deterioration is a result of an outdating, that is, the channel has changed from the
last estimation. This type of degradation is independent of the operating SNR, and
relates to the coherent time. The other scenario corresponds to a σe2 that increases with
SNR. This scenario relates the estimation quality to the operating SNR, as is usually
the case in reality. In the following computation we treat σe2 as a constant parameter,
irrespective of the aforementioned scenarios.
The capacity of a linear MIMO system with noisy channel estimation is not known.
Only upper and lower bounds on the mutual information are given in literature [35].
A lower bound on the mutual information is computed choosing a Gaussian input
distribution and the entropy of a Gaussian random variable, whose variance is given
x
by the mean square error of the MMSE estimate of s given y and H:
1 x xH
Ilower
x
s; y H EH
x log INR
SNR
2
1 σe SNR NT
HH , (3.16)
where we used that E rsH ss NPT I and the SNR definition from (3.2). An upper bound
on the mutual information is computed using the fact that the Gaussian distribution
maximizes entropy. It can be written as a function of the lower bound as follows
σe2 SNR
Iupper
x
s; y H Ilower
x
s; y H NR E s log
1
1 σe2 }s}2
. (3.17)
However, these bounds are tight only for small values of σe2 , and are not suited as
reference curves for larger values of σe2 (cf. Sec. 3.3). For a better comparison with our
nonlinear MIMO achievable rates, we will numerically compute the achievable rate of
the linear reference system with a Gaussian input distribution. This computation is
very similar to the nonlinear MIMO achievable rate computation that follows in the
next session, and we will describe both in parallel.
20
3.2 Rates with Noisy Channel State Information
x
We compute the mutual information between s and y given a channel estimate H:
x
I s; y H x
x I s; y H
EH p
H
EH
p
x h y H h p
y s, H . (3.18)
y p
g Hs r
w gprq, (3.20)
r
where w w Es. Here we assume that w and E are statistically independent. This
is a valid assumption since coding is performed across different channel realizations,
which is common when computing ergodic rates. Hence, for a given realization of s s,
wr is Gaussian distributed, as
¸
NT
r |s
w CN 0, σw2 σe2 |sj |2 I . (3.21)
j 1
x H,
Furthermore, for a given realization of H p r is also Gaussian distributed, as
¸
NT
p
r s, H CN p
Hs, σw2 σe2 |sj |2 I . (3.22)
j 1
We use the above distribution for the computation of all required entropies. In the
case of the linear MIMO system, (3.22) corresponds to the distribution of y |s, H,
p since
21
Chapter 3 Information Theoretic Limits
averaged over many realizations of s. Note that (3.23) is independent of the channel
estimate H,p which affects only the mean of the distribution in (3.22). This is true only
for the linear MIMO case.
NR
¹
f y s x
s, H p
H p
f yi eT
i Hs , (3.24)
i 1
using the fact that the composite noise w r remains white, i.e., the correlation matrix
is diagonal. The distribution f pyi |ei Hsq per receive antenna is then given by (2.6)
Tp
°
and (2.7), evaluated by setting xi eT
i Hs and σw σw
p 2 2
j 1 |sj | in those
σe2 N T 2
equations, using (3.22). Based on the above equation, we compute the remaining
entropies numerically.
for a given channel estimate. The conditional distribution in (3.25) is described below
(3.24). This concludes the computation of hpy |s, Hp q both for linear (see (3.23)) and
nonlinear MIMO.
22
3.3 Numerical Evaluation
Finally, hpy |H
p q is computed in the same way for linear and nonlinear MIMO. How-
ever, unlike in the perfect CSI case, the distribution f py |H
xq is not known analytically
even for a SISO system. This is due to the term (Es) in (3.19), which consists of the
sum of products of Gaussian r.v.’s. Hence we resort to MC estimation of f py |Hxq as
follows. First we create N samples y yi , generated according to f py |H
xq. That is,
p the source of uncertainty are s, E and w. The entropy
for a fixed channel estimate H,
hpy |H
xH p q is computed with MC integration
1 ¸
N
x
h y H p
H N log f yi H H .
x p (3.26)
i1
»
1 ¸
M
x
f yi H Hp f psq f yi s, H
xH
p ds
M j 1
f yi s sj , H
xH
p .
In this section we present numerically computed achievable rates for various config-
urations. In all following plots, solid, dotted and dashed lines stand for linear, am-
plitude and phase detection, respectively. The input s is distributed according to
s CN p0, σs2 I q, unless stated otherwise. The SNR is defined in (3.2). The chan-
nel matrix entries are modeled as i.i.d. zero-mean unit-variance circularly symmetric
Gaussian random variables. Although we do not explicitly simulate them, we expect
that rank-deficient channel matrices reduce the achievable rates of nonlinear MIMO
systems in a way similar to linear MIMO systems.
23
Chapter 3 Information Theoretic Limits
10
Amplitude Detection Amplitdue Detection, 1x1
Phase Detection, Gaussian input 25 Amplitdue Detection, 2x2
9
Phase Detection, Uniform phase input Amplitdue Detection, 3x3
8 Linear Detection Phase Detection, 1x1
Rate (bits/complex dimension)
4
10
3
2 5
0 0
−10 −5 0 5 10 15 20 25 30 −10 −5 0 5 10 15 20 25 30
Average SNR (dB) Average SNR (dB)
(a) (b)
Figure 3.1: (a) SISO system, perfect CSI, (b) Perfect CSI, NT NR t1, 2, 3u.
First we consider the case of perfect CSI at the receiver, as described in Section 3.1.
Fig. 3.1a depicts the rates for a SISO system, computed by averaging (3.5) over many
channel realizations, using (3.7), (2.6), (2.7) and the analytic entropies for amplitude
and phase detection (known in the SISO case only). Nonlinear detection is clearly infe-
rior to linear due to the non-invertible, information-destroying processing by function
g pq. Furthermore, the slope of the rate curve for nonlinear detection at high SNR is
half the slope of linear detection, and thus, equal to the slope of linear detection using
one-dimensional (real-valued) signals. Appendix B includes an approximate analysis
showing that the pre-log factor for amplitude and phase detection is indeed 12 . We also
plot the rate for phase detection when the input signal has unit amplitude and uniform
phase. This distribution yields higher rate, although it is not capacity achieving, con-
trary to what is stated in [36]. For higher number of receive antennas, simulations show
that the uniform distribution yields the same rates as the complex Gaussian distribu-
tion. This effect rises from the MIMO channel multiplexing. In the rest of this section
we use only the complex Gaussian input distribution. An observation which will also
hold in the following plots is that phase detection outperforms amplitude detection
at all SNR. In other words, phase conveys more information than amplitude. This is
due to the fact that the conditional pdf in (2.7) depends on the amplitude zi through
ρi , and hence carries information regarding the strength of the channel. On the other
side the pdf of the amplitude in (2.6) is totally independent of the phase. This consti-
24
3.3 Numerical Evaluation
4.5
20 Amp. Det., 2x2
Amp. Det., 2x3
Amp. Det., 2x4 4
18 Phase Det., 2x2
Phase Det., 2x3
16
8 2
6
1.5
4
1 Amplitude detection
2 Phase detection
Linear detection
0 0.5
0 5 10 15 20 25 30 1 2 3 4
Average SNR (dB) Number of receive antennas
(a) (b)
Figure 3.2: NT 2 and varying NR , perfect CSI. (a) Achievable rates and (b) spatial
multiplexing gain in real-valued dimensions.
tutes a fundamental difference between amplitude and phase detection, which we will
encounter in the following chapters as well.
N N MIMO: Fig. 3.1b depicts the achievable rates of MIMO system with NT
NR . As expected, the slope of the nonlinear rate curves increases when increasing the
number of transmit and receive antennas. Furthermore, the slope of the nonlinear rates
remains approximately half the slope of the respective linear capacity, like in the SISO
case. For example, for NT NR 2, the slope of amplitude and phase detection at
high SNR is the same as the slope of linear detection with NT NR 1. Hence, we
conjecture that an N N nonlinear MIMO system achieves N {2 spatial multiplexing
gain.
2NR MIMO: Fig. 3.2(a) compares a system with NT 2 and variable number of
receive antennas NR 1, . . . , 4. For linear detection, the slope of the capacity curves
for NR ¥ 2 remains the same, since the degrees of freedom Nmin 2 are unchanged
and fully exploited. The parallel shift is the result of a power gain. However, the slope
of the nonlinear rate curves continues to increase as we add more receive antennas, but
always remains lower than the linear detection slope. We quantify this observation by
approximating the rate curves with R α logp1 βSNRq and plot twice the estimated
spatial multiplexing gain α in Fig. 3.2(b) (this corresponds to real degrees of freedom).
We see that the nonlinear detectors exploit the additional receive antennas to achieve
higher spatial multiplexing gains. We conjecture that they always miss one real degree
of freedom, i.e., they reach a spatial multiplexing gain up to 2Nmin 1 3 in real
25
Chapter 3 Information Theoretic Limits
12
Amplitude Detection, 2x2
Amplitude Detection, 2x3
Phase Detection, 2x2
10 Phase Detection, 2x3
Linear Detection, 1x1
0
−10 −5 0 5 10 15 20 25 30
Average SNR (dB)
Figure 3.3: Linear SISO versus nonlinear MIMO rates, perfect CSI.
Before proceeding to the noisy CSI achievable rates, we take a look at the outage
probability of nonlinear MIMO systems in a slow fading scenario, where averaging over
different channel realizations is not possible. The outage probability is non-zero for
26
3.3 Numerical Evaluation
0 0
10 10
−1 −1
10 10
Outage Probability
Outage Probability
−2 −2
10 10
−3 −3
10 10
Amplitude Det. ,R=1 Phase Det., R=1
Amplitude Det., R=2 Phase Det., R=2
Amplitude Det., R=3 Phase Det., R=3
−4 Amplitude Det., R=4 −4 Phase Det., R=4
10 Linear Det., R=4 10 Linear Det., R=4
t u Linear Det., R=6
Linear Det., R=8 t u Linear Det., R=6
Linear Det., R=8
Linear Det., R=12 Linear Det., R=12
5 10 15 20 25 30 0 5 10 15 20 25 30
Average SNR (dB) Average SNR (dB)
(a) (b)
Figure 3.4: Outage probabilities for different outage rates, (a) amplitude and linear detec-
tion, (b) phase and linear detection. NT NR 2.
any given outage rate. We consider a NT NR 2 system and investigate the outage
probability for different outage rates. The slope of the outage probability curves at
high SNR reflects the diversity that is achieved at a given outage rate. It is known
that the maximum diversity with linear detection is NT NR [3]. However, it is not clear
whether nonlinear detection achieves the same maximum diversity.
Fig. 3.4a depicts the outage probability curves for amplitude and linear detection.
For producing these curves we have simulated many channel realization for given SNR
values, and computed the fraction of cases where the achievable rate was less than
the target outage rate. The leftmost curves drop four orders of magnitude within a
range of 10 dB’s, which shows that amplitude detection also achieves the maximum
diversity of the MIMO channel. The same holds for phase detection, as illustrated
in Fig. 3.4b. We also observe that the slope of the outage probability curves flattens
for larger outage rates. This is due to transition toward other operating regions on
the diversity-multiplexing tradeoff plane, as explained in [37]. These results hint that
nonlinear detection, like linear detection, exploits the full MIMO diversity.
We proceed to achievable rates with noisy channel state information. First, we consider
a scenario where the channel estimate degradation is due to outdated information, i.e.,
27
Chapter 3 Information Theoretic Limits
10 10
Linear Detection (Bounds) Linear Detection
9 Linear Detection (Numerical) 9 Phase Detection
Amplitude Detection
8 8
Rate (Bits/complex dimension)
6 6
σe2 t0, 0.001, 0.01, 0.1, 0.5u
t0, 0.001, 0.01, 0.1, 0.5u
5 5
σe2
4 4
3 3
2 2
1 1
0 0
−10 −5 0 5 10 15 20 25 30 −10 −5 0 5 10 15 20 25 30
Average SNR (dB) Average SNR (dB)
(a) (b)
Figure 3.5: (a) SISO system, noisy CSI. Amplitude Detection, (b) SISO system, noisy CSI.
Phase Detection.
σe2 is constant for all SNR. The achievable rates of the linear reference system are
computed as described in Sections 3.2.1 and 3.2.2. We depict the lower and upper
bounds of (3.16) and (3.17) only in Fig. 3.5a to illustrate that the numerically computed
curves are more useful than the two bounds, especially at high σe2 . The estimation error
variance is chosen as σe2 t0.5, 0.1, 0.01, 0.001u and corresponds to an estimation SNR
of t3, 10, 20, 30u dB, respectively, while σe2 0 corresponds to perfect CSI. Curves with
the same marker correspond to the same value of σe2 .
SISO: Figures 3.5a and 3.5b depict the rates for a SISO system for amplitude and
phase detection, respectively. Similar to the linear reference system, the achievable
rates reach a threshold for high SNR. This threshold is reached sooner for large esti-
mation error variances. The rate loss at σe2 0.001, which implies an excellent channel
estimate, is very low. Phase detection remains always better than amplitude detection
for the same σe2 , similar to the perfect channel knowledge case. This result holds for
various antenna configurations, pointing out that both techniques are equally robust
to channel estimation errors.
2 2 MIMO: Fig. 3.6a depicts the achievable rates for a NT NR 2 MIMO system.
The behavior is similar to the SISO case. Phase detection performs well with respect to
linear detection at low SNR for perfect CSI. Now we can see that phase detection can
even outperform linear detection, when the channel estimation quality of the former
is somewhat better than that of the latter, e.g., σe2 0.01 for phase detection and
σe2 0.1 for linear detection. Concluding, we can say that both amplitude and phase
28
3.4 Fundamental Characteristics of Nonlinear MIMO
18 18
Linear Detection Linear Detection
Amplitude Detection Amplitude Detection
16 16
Phase Detection Phase Detection
14 14
12 12
10 10
6 6
4 t u 4
2 2
0 0
−10 −5 0 5 10 15 20 25 30 5 10 15 20 25 30
Average SNR (dB) Average SNR (dB)
(a) (b)
Figure 3.6: (a) 2 2 MIMO system, noisy CSI, (b) 2 2 MIMO system, noisy CSI with
SNRest SNR.
detection exhibit the same behavior as linear detection when the channel estimation is
not perfect.
Finally, we simulate the achievable rates when the estimation SNRest 10 log10 σe2 is
a function of the operating SNR. This scenario is closer to reality, since the SNR during
the training and data transmission phases is usually equal. In this case, operating at
high SNR implies good quality channel estimation. We consider the case where SNRest
is equal to the operating SNR. The results are depicted in Fig. 3.6b for a 2 2 MIMO
system, along with the case of perfect CSI (marked with a circle). As can be seen,
the rate curves are parallel to the perfect CSI curves. The rates do not saturate for
increasing SNR, since the channel estimation quality also increases with the SNR.
In the previous sections we saw that nonlinear MIMO systems also exploit the spatial
degrees of freedom, albeit in a different way than linear MIMO systems. Although phase
and amplitude detection differ a lot, both nonlinear detectors achieve half the maximum
degrees of freedom of a N N MIMO system. Adding more receive antennas, more
dimensions can be exploited—a result that is not intuitively straightforward. Both
29
Chapter 3 Information Theoretic Limits
nonlinear receivers exhibit a similar robustness to noisy CSI as their linear counterparts.
Compared to a nonlinear SISO system, which they intend to substitute, nonlinear
MIMO systems achieve significantly higher rates. In the following section we discuss
how this rate gain affects wireless sensor networks.
In this section we discuss the impact of using a nonlinear MIMO receiver in a wireless
sensor network, taking the above information theoretic results into account. There are
certain properties that are critical for the power consumption of the radio system on
a sensor node. A sensor network is characterized by sporadic transmission (low duty-
cycle) of short packets over small distances. This implies that the output power of the
transmitter is relatively small, and does not dominate the transmit power consumption.
Furthermore, the radio system must be waken-up every time a transmission is pending,
which means that the power consumed during the start-up phase is very important.
However, since the transmission time is short due to the small packet size, the start-up
power consumption is equally important to the actual transmission. Hence, increasing
the transmission rate alone is meaningful only if accompanied with a short start-up
time, as elaborated in [10] and [38]. Concluding, the decisive factors for a low-power
sensor node are low transmit/receive power consumption, through suitable choice of
(nonlinear) modulation, short start-up time and high rate. Our proposed nonlinear
MIMO systems possess all these properties.
As we saw in Section 3.3, nonlinear MIMO systems exploit spatial degrees of freedom
and achieve higher rates than nonlinear SISO systems. Compared to a linear SISO sys-
tem (cf. Fig. 3.3), the achievable rate of amplitude and phase detection are similar (or
even better for phase detection) and the scaling with SNR (spatial multiplexing gain)
is at least as good as that for linear SISO. However, nonlinear MIMO implementations
have not been considered yet. As explained in Section 1.2, facilitating multiple anten-
nas on a node is possible with the use of compact MIMO-antennas. Regarding the RF
circuitry, a nonlinear MIMO receiver would use multiple transmit/receive chains of a
nonlinear SISO design, e.g. [9] for amplitude detection. Associated properties like very
low start-up time and reduced complexity will remain the same, despite the multiple
antennas. The analog front-end power consumption will be increased, but still remain
low compared to a linear I/Q receiver, since the low-power nonlinear SISO designs are
orders of magnitude more efficient (cf. [9], [12]). Thus, nonlinear MIMO nodes would
30
3.4 Fundamental Characteristics of Nonlinear MIMO
improve the power consumption in a sensor network without requiring any changes in
the higher-layer network protocols.
31
Chapter 3 Information Theoretic Limits
32
Chapter 4
In this chapter we take a step away from the information theoretical considerations of
Chapter 3, and consider nonlinear MIMO systems with practical modulation alpha-
bets. We investigate possible modulation schemes for amplitude and phase detection
in conjunction with the ML detector for these systems. In fact, the ML detector dic-
tates what modulation is possible for each system. We start our analysis revising the
SISO performance of amplitude and phase detection, and then proceed to the multiple
antennas extension. In Section 4.3 we investigate how the diversity of the nonlin-
ear MIMO ML detector is affected for modulations that perform spatial multiplexing.
This behavior of nonlinear MIMO systems is unique and not known from linear MIMO
systems.
For this first step towards a more realistic system, we assume that the channel is
perfectly known at the receiver throughout this chapter. Later on, in Chapter 5 we will
investigate channel estimation schemes for nonlinear MIMO systems, and observe that
complete knowledge of all complex-valued channel coefficients is not always required.
Finally, our analysis of nonlinear SISO systems in Sections 4.2.1 and 4.2.2 differs from
common practice in literature, since we analyze modulations that are typically detected
without channel knowledge (non-coherent), by assuming that perfect CSI (coherent) is
available. Sections 4.3.1 and 4.4.1 have been published in [21].
33
Chapter 4 Modulation and Maximum Likelihood Detection
4.1 Modulation
In Chapter 3 we considered almost exclusively complex Gaussian modulation alphabets.
This choice was motivated by the capacity achieving distribution for linear systems,
and the fact that the Gaussian distribution aids mathematical analysis. For SISO
phase detection, we also used a constant amplitude–uniform phase distribution that
yielded higher rate compared to the Gaussian distribution (cf. Fig. 3.1a). In that
case, simulations showed that the phase detector extracts more information (higher
rate) when it has to detect between symbols with different phases (uniform phase,
constant amplitude), instead of symbols with different phase and amplitudes (Gaussian
distribution). However, in this and the following chapters we will deal with uncoded
transmission, and hence we no longer consider Gaussian alphabets.
Intuitively, an amplitude detector can only distinguish between different ampli-
tude/energies of symbols, since the phase of the symbols is removed at the receiver.
Analogously, a phase detector can only detect between symbols with different phases,
since the amplitude/envelope of the symbols is discarded. These observations are al-
most correct. In fact, the form of the ML detector rule dictates what symbols can or
can not be distinguished/detected by the amplitude or phase detector. Hence, we will
examine modulation parallel to the ML detector in the next section.
pps|y q, s P S, (4.1)
where S is the modulation alphabet. The MAP detector is based on the selection of
the signal that yields the maximum a posteriori probability (4.1). From Bayes rule we
have that
f py |sq ppsq
pps|z q
f py q
, (4.2)
where f py |sq is the probability density that a specific symbol s will lead to the obser-
vation y. If we assume that the a priori probability ppsq, that a specific s was sent, is
34
4.2 Maximum Likelihood Detection
constant for any s, i.e. s is uniformly distributed, the computation of the a posteri-
ori probabilities depends only on the conditional densities f py |sq, since the probability
density f py q is fixed for a given s. The ML criterion is based on the selection of the
signal that yields the maximum conditioned density f py |sq. Consequently, for uniform
s, we have that
y |hs w | |z w |. (4.5)
2y y2σ2|z|2 2y |z |
f py |z q e w I0 . (4.6)
σw2 σw2
As we see, the pdf depends only on the norm of z hs, meaning that any symbols with
the same amplitude and different phase will be indistinguishable, since they yield the
same pdf. On the other side, symbols with different amplitudes lead to different pdfs
and are thus detectable. A straightforward choice is binary OOK, where the symbols
consist of s t0, Au, A P R . Multilevel M On-Off Keying (M-OOK) can also be
35
Chapter 4 Modulation and Maximum Likelihood Detection
2
p(y|s=0)
1.8 p(y|s=A)
Threshold b
1.6
1.4
1.2
pdf 1
0.8
0.6
0.4
0.2
0
0 0.5 b 1 1.5 2 2.5 3
y
?
Figure 4.1: Computation of error probability, A 2
2, σw 0.2.
used, since different levels can be separated by the amplitude detector. At this point
we note that Pulse-Amplitude Modulation (PAM) (cf. [39]) can not be readily used,
and the name is hence misleading. In fact, PAM uses both positive and negative levels,
which means that the phase of the symbols is also modulated. An amplitude detector
would only differentiate between half the PAM signals, those that differ in the absolute
value of the amplitude. Binary PAM is identical to Binary Phase-Shift Keying (BPSK)
and can as well not be detected.
Let us now consider simple OOK and ML detection. The signals s t0, Au are
transmitted with equal probability over a fading channel h. The observed signal is
either of the two: #
y |w| if s 0,
? , (4.7)
|Ah w| if s A 2Eb
where Eb is the average bit energy. We will compute the average
Bit Error Ratio (BER). First, let us assume the channel is fixed. This corresponds
to an Additive White Gaussian Noise (AWGN) channel. This detection problem has
been studied extensively in literature. Nevertheless, as we will see there is no analytical
solution. The pdfs for the two hypotheses are:
2y σy22
f py |s 0q e w (Rayleigh distribution), (4.8)
σw2
2y y2 σ|2Ah|2 2y |Ah|
f py |s Aq e w I0 (Rice distribution). (4.9)
σw2 σw2
The probabilities of error are not equal for the two hypotheses. They are given by
36
4.2 Maximum Likelihood Detection
f py
b|s A
q f py b|s 0q (4.12)
| |
Ah 2
2b|Ah|
e 2
σw
I0 1 (4.13)
?
σw2
2b|h| 2γ
e2|h| γ .
2
I0 (4.14)
σ w
This threshold is tight for very low and very high SNR. Fig 4.2a depicts the exact
threshold, computed numerically, and the simple threshold given by the above approx-
imation. We see that the approximation is very good, and tight for low and high SNR.
As the SNR increases, the threshold becomes asymptotically equal to
A|h|
lim b 12 p0 A|h|q, (4.16)
γ Ñ8 2
Proceeding, since the two symbols are equiprobable, the average BER for a given
1
Since the problem is in essence real-valued, the SNR is usually defined using only the real noise
variance σw 2
{2 in literature. ?
2
We use (7-4-13) from [40] and substitute b0 Ñ σw2b and γ Ñ 2|h|2γ. The difference to the analysis
in [40] is the noise variance and the fading channel.
37
Chapter 4 Modulation and Maximum Likelihood Detection
3.5 0
Exact Threshold 10
Approximate Threshold
High SNR asymptotic
3
−5
10
Detection Threshold b
2.5
−10
10
BER
2
−15
10
1.5
−20
10
1
BER with threshold b
High SNR approximation
−25
BER with exact threshold
0.5 10
−10 −5 0 5 10 15 20 25 −10 −5 0 5 10 15 20
SNR SNR
(a) (b)
Figure 4.2: (a) Detection thresholds for OOK, (b) BER in AWGN, |h| 1.
channel h is given by
pPeS PeM q. Prperror|hq
1
(4.17)
2
An approximation to the BER for high SNR can be used employing an approximation
of the modified Bessel function of zeroth order for large arguments, and is given by [41]
Pe phq
1 |h2|2 γ .
p4π|h|2γ q1{4 e (4.18)
Fig. 4.2b shows the BER for the AWGN channel (h 1) using the approximate
threshold (4.15) and (4.10), (4.11), (4.17), as well as the high SNR approximation
(4.18). The BER with the numerically computed ‘exact’ threshold is also plotted, and
shows that the approximate threshold (4.15) is in fact a great approximation.
Averaging (4.18) over all realizations of the channel coefficient, we obtain an approx-
imation of the average BER for high SNR
»8 1{4
|h|2 γ 2
Γ 34
Pse E hrPephqs f p|h|q
p4π|h|2γ q1{4 e d|h| pγ p2
1
2 π
γ q3 q1{4
, (4.19)
0
where |h| is Rayleigh distributed with σ 2 σh2 {2 and σh2 1. The Taylor series
expansion of (4.19) around γ1 0, that is, for high SNR, yields
2 1 4 { 3
Γ
Pse π
SNR
4
O
1
SNR2
(4.20)
38
4.2 Maximum Likelihood Detection
which shows that OOK in fast fading has diversity 1. We also verify this through
simulation. The fact that OOK has diversity 1 means that the dimensionality loss of
amplitude detection has no impact on the diversity, but only induces an SNR penalty
when compared to coherent BPSK.
This analysis holds for binary OOK. As we will see later in Section 4.3.1, detection
of multilevel M-OOK has significant differences. Finally, note that optimum detection
of OOK requires computation of the two probabilities (4.10) and (4.11), and finally
comparison. Even using the approximate threshold from (4.15), update of the thresh-
old is required when the SNR changes. This presents an inconvenience of amplitude
detection that needs to be accounted for in practical implementations.
Assume, as above, transmitting s over the fading channel h CN p0, σh2 q. After the
addition of AWGN at the receiver, we observe the phase of the received signal
where ρ |z |2{σw2 and ∆φ rpy =pz qq mod 2π s P pπ, π s. The pdf depends both on
the phase and the absolute value of z, through ρ. Hence, in general, the phase detector
requires also knowledge of the norm (strength) of the channel h, or else, requires full
knowledge of the complex-valued channel coefficient. Now, assume we transmit one of
two signals s ts1 , s2 u. The two possible observed signals are:
#
y =phs1 w q if s s1 ,
. (4.23)
=phs2 w q if s s2
Figures 4.3a and 4.3b depict the conditional pdf for varying |hsi | and =phsi q 0, and,
varying =phsi q and keeping |hsi | constant, respectively. First, we observe that the
phase detector can distinguish between signals bearing the same phase, and different
amplitudes: =ps1 q =ps2 q (cf. Fig. 4.3a). Hence, the phase detector can even surpris-
ingly detect OOK signals, or any Quadrature-Amplitude Modulation (QAM) scheme
39
Chapter 4 Modulation and Maximum Likelihood Detection
1.8 1.8
x=1
1.6 x=0.75 1.6
x=0.25
1.4 x=0.025 1.4
1.2 1.2
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
−3 −2 −1 0 1 2 3 −3 −2 −π/2 −1 0 1 π/2 2 3
y y
(a) (b)
Figure 4.3: (a) Pdf’s for fixed =phsi q 0, varying x |hsi|, (b) Pdf’s for fixed x |hsi|,
=phs1 q π{2, =phs2 q π{2.
sp j
y » 0 (4.24)
j
sp
3
We use this definition for BPSK rather than the popular s t1u, in order to avoid the modulo
wrapping f py |z q for s 1 in Fig. 4.3b.
40
4.2 Maximum Likelihood Detection
computed for every phase φi =phsiq. The above metric considers the wrapping
of the distance computation beyond the points π in Fig.4.3b. Analyzing the BER
using (4.22) and (4.24) or (4.25) is cumbersome. However, since the phase detector is
optimum for the detection of BPSK, we can use the linear detector instead. Hence,
the BER is given by [3]
a
Pe phq Q 2|h|2SNR (4.26)
as a function of the channel realization h, and SNR Eb {σw2 . Averaging over the fading
channel with σh2 1 we obtain the average BER
Pse E rPephqs 1
2
1
SNR
1 SNR
. (4.27)
The Taylor series expansion of (4.27) at high SNR yields (see [3])
Pse 1
4SNR
O
1
SNR2
(4.28)
which indicates that ML detection of BPSK in a fading channel has diversity 1. This
result carries on to our phase detector as well.
Finally, we show the BER of SISO amplitude and phase detection in Fig. 4.4. The
performance of BPSK detection is identical using the metric of (4.22), the metric in
(4.24), the linear BPSK detector or the analytic expression in (4.27). The BER of OOK
phase detection is also depicted. Detection is possible, but performance is very poor.
Finally, the average BER of OOK amplitude detection is also shown for comparison.
We observe that the amplitude detection is poor compared to phase detection of BPSK.
This is due to the non-optimality of OOK. Closing, all schemes (except phase detection
of OOK) have diversity order 1 in a fading channel.
41
Chapter 4 Modulation and Maximum Likelihood Detection
0
10
−1
10
Average BER
−2
10
−3 Amplitude Det.,OOK
10 Phase Det., metric
Phase Det., Euclidean dist.
Phase Det., linear recever
Phase Det., analytic
−4
Phase Detection, OOK
10
−10 −5 0 5 10 15 20 25 30
SNR
Figure 4.4: Comparison of average BER’s for SISO amplitude and phase detection.
Now we extend to MIMO amplitude and phase detection. The ML detector from (4.4)
can be written as
¹
NR
ŝML argmax f pyi |zi q (4.29)
s PS i 1
using (2.5). The conditional pdfs on every receive antenna are given by (2.6) and (2.7)
for amplitude and phase detection, respectively.
Amplitude Detection In the case of amplitude detection, the observations are similar
to the SISO case analyzed above. Indeed, only the amplitude of the transmitted symbol
can carry information. When the transmitter has no knowledge of the channel, no
precoding is possible in order to construct specific amplitudes at the receiver. Hence,
transmitting real-valued signals that differ only in amplitude is the only choice. We
extend OOK modulation to a MIMO system as follows:
42
4.3 Spatial Multiplexing and Diversity
Phase Detection In Section 4.2.2 we saw that the phase detector is able to detect
any complex-valued modulation, but, phase-only modulations, like M-PSK, are more
efficient. The latter leads additionally to very efficient detection, since the simple
distance metric (4.25) suffices for detection. Finally, knowledge of the channel norm |h|
is not required for M-PSK, and the conditional pdfs for all symbols are shifted version
of the same pdf. However, the situation in MIMO is different. The multivariate phase
distribution in (4.29) is composed of different pdfs in every dimension. This occurs
because of the different MIMO channel coefficients. Hence, it is not possible to come
up with a simple decision rule, like in SISO M-PSK. Similar to amplitude detection,
computation of the exact conditional pdf is required for all possible symbols.
In analogy to Ex-OOK, we define Extended Binary Phase-Shift Keying (Ex-BPSK)
as our reference modulation for MIMO phase detection. The modulation symbols have
the form
sPS tα1uN T
, (4.31)
a
where α1 Es {NT and Es is the average symbol energy. The uncoded rate is NT
bits/s/Hz and is achieved through spatial multiplexing. In Section 4.3.2 we investigate
how the use of spatial multiplexing affects ML detection and diversity of MIMO phase
detection.
Modulation alphabets that perform spatial multiplexing, like Ex-OOK and Ex-BPSK,
have an impact on the diversity order of MIMO ML amplitude and phase detection. In
the following we investigate the Pairwise Error Probability (PEP) between two equally
likely symbols and discover the mechanism that differentiates MIMO from SISO am-
plitude and phase detection.
43
Chapter 4 Modulation and Maximum Likelihood Detection
For the time being we do not define the structure of the symbols. The corresponding
received signal on the ith antenna is
¸NT
yiY
j 1
hij sY
j
wi
ziY wi , Y P tA, Bu. (4.32)
f yi |s s Y
e w I0 . (4.33)
σw2 σw2
where wi1 wi ej =pziq . We stack all wi1 , i 1, . . . , NR in vector w 1 CN t0, σw2 I u,
that has the same noise statistics as w. Observe that (4.32) and (4.34) are equivalent
detection problems with respect to the ML detector. Eq. (4.34) is another manifestation
that the phase of the received signal carries no information. We will consider detection
problem (4.34) in the following.
The distribution of ziY given sY is Gaussian ziY CN t0, σh2 EsY u, i. Furthermore,
ziY is independent across different antennas
E ziY zjY, δi,j σh2 EsY (4.35)
for the same symbol Y. In (4.35) we make use of the fact that vecphij q CN p0, σh2 I q.
This is in general not true among different symbols at the ith receive antenna
¸
NT
E ziA ziB, σh2 sA B,
(4.36)
j sj ,
j 1
44
4.3 Spatial Multiplexing and Diversity
Since the noise is independent between different receive antennas, the distribution
of f py |z Y q can be factored like
¹
NR
f y |z Y f yi |ziY . (4.37)
i 1
Consequently, the log-likelihood ratio can be written as
f y |z A ¸
NR
f yi |ziA
LLR ln ln f py |zBq ,
f py |z B q
(4.38)
i1 i i
which is the sum of the Log-Likelihood Ratios (LLRs) at each receive antenna. We
will treat the detection problem at each antenna separately. The ML detector will
optimally combine the likelihood ratios at each antenna before taking the final decision.
The performance of the vector detection problem is closely connected to the detection
problem at each antenna.
Let us now consider the scalar detection problem at the ith receive antenna. We
distinguish between two cases: either ziY 0, when sY 0, or ziY 0 when at least
one entry of sY is non-zero. This separation leads to two distinct scalar detection
problems4 :
#
yiA |wi1 | , sA 0,
D1i : (4.39)
yiB |ziB | wi1 , sB 0, ziB 0
#
D2i :
yiA |ziA| wi1 , sA
0, ziA 0 (4.40)
yiB |ziB | w 1 , sB
i 0, ziB 0
where ziA ziB . The receiver has to decide between a Rayleigh and a Rice distri-
bution for the two hypotheses in D1i , and between two Rice distributions in D2i . The
detection problem D1i has been studied in Section 4.2.1 and corresponds to detection of
OOK. D2i corresponds to the PEP of two non-zero symbols in M-OOK. An analytical
solution to D2i does not exist, so we will use an auxiliary detection problem Li2 whose
error probability is a lower bound to D2i .
Diversity order of D1i The detection problem D1i has been studied in Section 4.2.1,
using slightly different notation and scalings. Use of the norm of ziB in (4.39) does not
matter, since problems (4.32) and (4.34) are equivalent, as we discussed earlier. From
4
We denote with D1 and D2 the respective vector detection problems.
45
PSfrag 4 Modulation and Maximum Likelihood Detection
Chapter
|ziY | yi 2
Li
|| D2i
yi
PseD1
i
π
γ
Γ 4
O
1
γ2
, (4.41)
Li
D2i is obtained from Li2 by taking the norm of yi 2 as depicted in Fig. 4.5. Since D2i is a
processed version of Li2 , the average probability of error of Li2 is always lower or equal
to that of D2i .
PseL2
i
¤ PseD . i
2 (4.43)
In fact processing will either lead to a sufficient statistic, and hence to identical error
performance, or will destroy information, and thus deteriorate error performance.
From (4.43) we observe that the diversity order of D2i can not exceed the diversity
order of Li2 , and, hence, Li2 provides an upper bound. We can perform the same lower
bounding to the error probability of D1i too, however this is not necessary since (4.41)
provides the exact diversity order.
46
4.3 Spatial Multiplexing and Diversity
can be computed using the mean between the two hypotheses as a threshold:
PeLi2
zsiA , zsiB
Pr yi | ziA | |ziB |
ss A
(4.44)
2
Pr wi1 ¡
| | |ziB|
zA
i
(4.45)
2
?
Q
zA
i
a
| | |ziB| Q Es |zsiA | |zsiB |
a (4.46)
2 σw2 {2 2 σw2 {2
Q zsiA | || | zsiB γ
2
. (4.47)
We used the variance of the real part of the noise in (4.46), since detection takes
place only in the real dimension. At high SNR, the Rice distributions in D2i can be
approximated by Gaussian distributions5 (cf. [42]) and the probability of error of D2i
is well approximated by (4.47). Hence, the bound is tight at high SNR. In this regime,
the detection threshold of D2i is the mean of |zsiA | and |zsiB |. Furthermore, the norm
|zsiY | is important, meaning that the distribution of the power of sY across the transmit
antennas does not change the probability of error, which is only affected by the total
power of sY .
The average probability of error is computed by averaging over zsiA and zsiB which
capture the fading channel effect
8
¼
Pse
Li2
f |zsiA |,|zsiB | |zsiA |, |zsiB | PeL2 zsiA , zsiB d|zsiA |d |zsiB |.
i
(4.48)
0
°NT
As we saw earlier, ziA and ziB are independent only if
A B,
j 1 sj sj 0. In this case,
their joint pdf is decomposed in the product of two Rayleigh distributions, and (4.48)
can be computed analytically in (4.49)
?8γ γ γ {
2 3 2
?4 γ
2 1
b
γ
p4 8 2γ q γ 6 44γγ ?8
? b
2γ
PseL2 ?
4 γ
? ? ,
2p2 γq 4 γp 2 4 γq
(4.49)
5
The Gaussian approximation is in this case tightly concentrated around the mean, and the negative
part of the distribution can be neglected.
47
Chapter 4 Modulation and Maximum Likelihood Detection
where we used σh2 1. The Taylor series expansion for high SNR yields
Ps L2
e
i
? 1 1{2 O
1
γ 3{2
. (4.50)
2γ
Thus, the diversity order of Li2 is only 21 , when ziA and ziB are independent. Otherwise,
the joint pdf is a correlated two dimensional Rayleigh distribution, which is known
analytically [31]. However, the average probability of error can not be computed an-
alytically anymore. Simulations show that the diversity order in this case remains 12 ,
meaning that correlation between ziA and ziB does not affect diversity.
Diversity and spatial multiplexing So far we considered the scalar detection problem
at each antenna separately. However, the ML decoder jointly detects the transmitted
signal using the whole received vector and (4.38). Unlike for linear MIMO receivers,
coming up with a combination scheme, e.g. maximum ratio combining, is not possible in
this case. This would significantly simplify the ML expression. However, we know that
the ML, or any combination scheme for that matter, will achieve at most a diversity
equal to the number of receive antennas (independent branches) times the diversity
order of the individual detection problems. This means, that for the vector detection
problem D1 , the ML detector will achieve a diversity of NR , while for D2 , the diversity
will not be larger than N2R . Notice that (4.38) is the same in the case of a 1 N receive
diversity scheme. The achievable diversity in that case equals N times the diversity of
the detection problem on each receive branch.
When it comes to evaluating the diversity of a modulation scheme with more than two
symbols, the poorest PEPs dominate the performance. As we saw, the fundamental
difference between problems D1 and D2 lies in the achieved diversity. However, D1
corresponds to sA 0, which means that the diversity of the PEP between any two
other symbols, not equal to 0, will be at most N2R . This occurs whenever we perform
spatial multiplexing, since in this case we will have at least 2NT symbols, and 2NT 1 of
them will not be equal to the all-zero symbol. Hence, performing spatial multiplexing
with a MIMO amplitude detection receiver reduces diversity to one half of the maximum
available diversity. The only way to achieve the maximum diversity is by constraining
the rate to only two symbols, sA 0 and sB 0, i.e. to a rate of 1 bit/s/Hz. In
this case, the multiple antennas at the transmitter are of no use, and a single-input
multiple-output system suffices.
Example: Let us consider Ex-OOK. We use 2NT symbols of the form
48
4.3 Spatial Multiplexing and Diversity
a
si P t0, 2Es {NT uNT . The uncoded rate of the modulation alphabet is NT bits/s/Hz,
achieved through spatial multiplexing. The PEP between the all-zero symbol and any
other symbol has the form of D1 , and diversity order equal to NR . However, the PEP
of all other combinations of symbols has the form of D2 . Those combinations where
° T A B,
at least one symbol has a zero for every j 1, . . . , NT , fulfill N
j 1 sj sj 0 and
the corresponding upper bound has been computed in L2 . The rest of the symbols,
a
where at least for one j both symbols equal 2Es {NT , do not fulfill the above con-
dition. In any case, the achievable diversity can not be higher than N2R . Since the
diversity loss occurs whenever we have more than 2 symbols, there is no modulation
alphabet with rate higher than 1 bit/s/Hz that can achieve full receive diversity. This
also applies to M-OOK SISO systems. The diversity loss effect that occurs due to D2
also affects M-OOK SISO systems: whenever we increase the rate above 1 bit/s/Hz
using MOOK, the detection between the non-zero amplitudes is of the form D2i , and
diversity drops to 21 .
Catastrophic channels: In the above analysis we assumed that ziA ziB in or-
der to be able to distinguish between the two symbols sA and sB . This is usu-
ally fulfilled when considering random channel coefficients. However, note that cer-
tain channel matrices that are considered as ‘ideal’ for linear MIMO systems, cause
problems in the case of amplitude detection. Consider for instance the case of a
Discrete Fourier Transformation (DFT) matrix. Such a matrix is full-rank and or-
thogonal (unitary). It is considered an ‘ideal’ channel matrix for linear MIMO systems,
since it enables all spatial streams (cf. (3.3)) and allows linear MMSE and zero-forcing
receivers to perform best. For amplitude detection however, since all entries of the
DFT matrix have unit amplitude, the condition ziA ziB is violated for all symbol
pairs that have equal number of non-zero entries. Hence, such a channel matrix would
lead to poor overall performance of the ML detector. In fact, the MIMO amplitude
detection receiver works best with channel coefficients that vary in amplitude, even
when the channel matrix is full-rank.
49
Chapter 4 Modulation and Maximum Likelihood Detection
the time being we do not define the structure of the symbols, but know already that
any sA sB can be distinguished by the phase detector. The received signal at the
ith antenna is
¸
NT
yiY = hij sY
j wi = ziY wi , Y P tA, Bu. (4.51)
j 1
The distribution of ziY given sY is Gaussian ziY CN t0, σh2 EsY u, i and (4.35), (4.36)
still hold. Again, we consider the detection problem at the ith receive antenna. We
distinguish between two cases: the two symbols are either a rotated version of each
other, or not. This separation leads to two distinct scalar detection problems:
#
D1i :
=
yiA ziA wi ,
sA 0, (4.52)
yi =
B
ziA ejφ wi , s B
sAejφ ñ ziB ziA ejφ, φ 0
#
yiA = ziA wi , sA ,
D2i : (4.53)
yiB = ziB wi , sB sA ejφ , |ziB | |ziA |, ziB ziA φ
In D1i , the two symbols have the same norm and differ only in the phase. If φ π,
this corresponds to BPSK. If φ π, it can be thought of as the detection between
two symbols with phase difference φ in M-PSK. This problem has been discussed
in Section 4.2.2. In D2i , we have a more general detection problem, where the two
pdfs differ in the norm, and eventually also in the phase. This detection problem has
fundamental differences compared to D1i .
Diversity order of D1i Problem D1i was analyzed in Section 4.2.2. The error proba-
bility was computed using the equivalent linear model, and
Pse 1
4SNR
O
1
SNR2
, (4.54)
which means the diversity of D1i equals 1. Note that contrary to M-OOK that had
diversity 21 (problem D2i in Section 4.3.1), in the case of phase detection M-PSK has
diversity 1. Hence, increasing the rate beyond 1 bit/s/Hz using multilevel modulation
does not affect the diversity of the ML phase detector.
Diversity order of D2i To facilitate understanding the difference between D1i and D2i
consider the example in Figures 4.6a and 4.6b. For this example assume NT 2. In
50
4.3 Spatial Multiplexing and Diversity
ziB p1q
ziA
ziB p2q ziB p4q
ziA
ziB p3q
ziB
(a) (b)
Figure 4.6: Examples of ziA and ziB on the complex plane, for (a) D1i , (b) D2i .
where h hi1 hi2 and h CN p0, 2σh2 q. The phase of the two hypotheses differs by
π. The channel phase =phq rotates the constellation, and different values of |h| lead to
different concentric circles in Fig. 4.6a. The two hypotheses lie on the circumference of
a circle, and their phase difference is always π, for any |h|. Different channel strengths
|h| scale the constellation, thus changing the distance between the two hypotheses.
The constellation collapses when a deep fade occurs: |h| ! 1, and the two hypotheses
coincide. Considering the pdfs on Fig. 4.3b, the channel fading scales both pdfs equally
without shifting them.
51
Chapter 4 Modulation and Maximum Likelihood Detection
which do not only differ by a phase shift. The receiver creates the two hypotheses:
$
¸
NT &z A hi1 hi2 h
ziY hij sY hi1sY1 hi2 sY ñ % iB , (4.58)
j 1
j 2
zi j phi1 hi2q jh1
where h hi1 hi2 and h1 hi1 hi2 . The two auxiliary variables are Gaussian
distributed h, h1 CN p0, 2σh2 q and statistically independent:
E rhh1 s E |hi1 |2 hi1 hi2 hi2 hi1 |hi2 |2 σh2 σh2 0. (4.59)
Fig. 4.6b depicts ziA and ziB on the complex plane, for different channel realizations.
Assume that h 1 ñ ziA 1. Different realization of h1 place ziB anywhere on
the complex plane. As a consequence, the phase difference between the two hypotheses
changes for every channel realization. The scaling effect of the signal space for different
channel realizations that we observed in D2i does not show up here. In fact, the signal
space changes completely between different realizations. Finally, a deep fade can occur
when h h1 , although none of the channels is weak. This corresponds to point ziB p4q
in Fig. 4.6b. Of course, deep fades also occur when |h|, |h1| ! 1. In terms of the pdfs,
every channel realization leads to pdfs that have arbitrary phase difference.
This behavior of detection problem D2i is reflected in the diversity, which, as in the
case of D2i for amplitude detection, is only 12 . Unfortunately we can not come up with
a tractable detection problem that is a lower bound to D2i , as we did for amplitude
detection. Instead we rely on simulation results which show that the diversity of D2i is
indeed 21 (cf. Section 4.4.2 and Fig. 4.9a).
Diversity and spatial multiplexing Using a similar reasoning as for the analysis for
MIMO amplitude detection, the ML detector will achieve a diversity of NR for the
NR
vector detection problem D1 , while for D2 , the diversity will be 2
.
When it comes to evaluating the diversity of a modulation scheme with more than
two symbols, the poorest PEPs dominate the performance. As long as the symbols
used differ only by a phase shift from each other, all the PEPs are detection problems
of type D1 . In this case, the overall diversity equals NR . However, such a modulation
does not require multiple transmit antennas, and boils down to M-PSK using a single
transmit antenna. Contrary to MIMO amplitude detection, the uncoded rate can in
this case be log2 M ¡ 1 bits/s/Hz and the ML detector will still achieve the full diversity
NR . However, if we perform spatial multiplexing, we are forced to use symbols that
52
4.4 Performance Results
are not phase shifted versions of each other. In this case, certain PEPs are detection
problems of type D2 , and the overall diversity is limited by these pairs and is only N2R .
Hence, similar to MIMO amplitude detection, spatial multiplexing modulation schemes
achieve only half the maximum receive diversity.
Example: Let us consider Ex-BPSK. We use 2NT symbols of the form
! a )N T
si P Es {NT . The uncoded rate of the modulation alphabet is NT bits/s/Hz,
achieved through spatial multiplexing. For every s P S, s also belongs to S. All these
pairs ts, su have PEPs of type D1 with diversity NR . However, the PEP between any
symbol s and s1 P S ztsu is a detection problem of type D2 , and the corresponding
diversity order is only N2R .
In this section we present BER results for amplitude detection MIMO receivers and
compare them to a linear MIMO receiver that uses uncoded BPSK and ML detection.
The reference system uses spatial multiplexing, hence transmitting NT bits/s/Hz, and
exploits the full receive diversity of the MIMO channel, equal to NR . For the am-
plitude detection MIMO system, we use two modulation schemes: OOK without spa-
tial multiplexing (rate R 1 bit/s/Hz) and Ex-OOK with spatial multiplexing (rate
R NT bits/s/Hz). The first modulation scheme consist of the symbols r0, . . . 0sT and
a a T
2Es {NT , . . . , 2Es {NT and is equivalent to using only one transmit antenna and
? (
the symbols 0, 2Es (Single Input Multiple Output (SIMO)). For a fair comparison
between systems with different rates we use the SNR per bit, given by Eb {N0 γ {R.
The channel has i.i.d. Gaussian entries with unit variance.
D2i and lower bound Li2 : Fig. 4.7a depicts the average error rate of the scalar de-
tection problem D2i (4.40) and the corresponding lower bound Li2 (4.42). The lower
53
Chapter 4 Modulation and Maximum Likelihood Detection
−1
10
−1 −2
10 10
BER
BER
−3 Linear BPSK, R=2 bits/symbol
10
Ampl. Detection, R=1 bit/symbol
Ampl. Detection, R=2 bits/symbol
PEP between [0 1]T and [1 0]T
scalar det. problem D2 PEP between [0 1]T and [1 1]T
−2
scalar det. problem L2 −4
PEP between [1 0]T and [1 1]T
10 10
−10 −5 0 5 10 15 20 25 30 −10 −5 0 5 10 15 20 25
Eb/N0 Eb/N0
(a) (b)
Figure 4.7: (a)BER of scalar D2i and Li2 , (b) BER plot for a 2 2 MIMO system.
bound exhibits slightly better error performance at low SNR, and the performance is
practically identical beyond 15 dB. The simulation shows that the diversity order of
D2i is in practice equal to that of the lower bound ( 21 ).
22 MIMO: Fig. 4.7b depicts the average BER for a 2 2 system. Amplitude
detection exhibits diversity equal to NR 2 when R 1 bit/s/Hz, and diversity equal
to N2R 1 when R 2 bits/s/Hz. In the case of spatial multiplexing, diversity is
destroyed by the pairs:
# +
sA , s B
P 0
,
1
,
0
,
1
,
1
,
1
,
1 0 1 1 0 1
which correspond to detection problems of type D2 . The first pair exhibits the property
that ziA and ziB are independent, while this is not the case for the other two pairs. The
average PEPs of these pairs are also shown in Fig. 4.7b. We see that correlation has no
impact on the diversity order. The small SNR gain of the second and third pair is due
to the higher transmitted energy of the respective symbols, compared to the first pair.
Furthermore, the upper bound to the diversity of D2 is actually achieved in practice.
In both cases, the performance is clearly inferior to the linear MIMO receiver, which
however requires higher power consumption and hardware complexity.
2NR MIMO: Fig. 4.8a depicts the average BER of a 2 3 and 2 4 system, with and
without spatial multiplexing. As predicted by the analysis of the PEPs, the diversity
order with spatial multiplexing is only N2R 1.5 and N2R 2, respectively. Without
54
4.4 Performance Results
−1 −1
10 10
−2 −2
10 10
Average BER
BER
−3 −3
10 10
Lin. BPSK 3x3, R=3 bits/sym
Ampl. Det. 3x3, R=1 bit/sym
Linear BPSK 2x3, R=2 bits/sym Ampl. Det. 3x3, R=2 bits/sym
−4 Ampl. Det. 2x3, R=2 bits/sym −4 Ampl. Det. 3x3, R=3 bits/sym
10 10
Ampl. Det. 2x3, R=1 bit/sym Lin. BPSK 3x4, R=3 bits/sym
Linear BPSK 2x4, R=2 bits/sym Ampl. Det. 3x4, R=1 bit/sym
Ampl. Det. 2x4, R=2 bits/sym Ampl. Det. 3x4, R=2 bits/sym
Ampl. Det. 2x4, R=1 bit/sym Ampl. Det. 3x4, R=3 bits/sym
−5
10
−10 −5 0 5 10 15 20 25 −10 −5 0 5 10 15 20 25
Eb/N0 Eb/N0
(a) (b)
Figure 4.8: (a) BER plot for a 2 3 and 2 4 MIMO system, (b) BER plot for a 3 3 and
3 4 MIMO system.
3NR MIMO: Fig. 4.8b finally depicts the same comparison for a 3 3 and 3 4 sys-
tem. This time, we have three levels of spatial multiplexing, yielding the uncoded rates
R 1, 2, 3 bits/s/Hz. The observation for the R 1 bits/s/Hz and R 3 bits/s/Hz
cases are as expected—the diversity order equals NR and N2R , respectively.! a The modu-
)
lation for the R 2 bits/s/Hz case has the form s ra1 , a1 , a2 s , ai P 0, 2Es {NT ,
T
such that two independent data streams are transmitted. The resulting diversity order
is again N2R . Note that the diversity loss occurs immediately when the rate increases
from 1 to 2 bits/s/Hz. However, further increasing the spatial multiplexing and the
rate has no impact on the diversity, which remains at N2R . Finally, note that the curves
for R 1 bit/s/Hz, for the 2 3 (2 4) and 3 3 (3 4) systems in Fig. 4.8a and
Fig. 4.8b, respectively, are identical. This is because all these systems are equivalent to
using a single transmit antenna, since no spatial multiplexing occurs, and hence, only
the number of receive antennas influences the performance.
55
Chapter 4 Modulation and Maximum Likelihood Detection
BER of D1
BER of D2
−1
−1
10
10
−2
10
Average BER
Average BER
−2
10
−3
10
(a) (b)
Figure 4.9: (a) BER plot for a 2 2 MIMO system, (b) BER plot for a 2 3 and 2 4
MIMO system.
In this section we present BER results for phase detection MIMO receivers and compare
them with the same linear MIMO receiver that uses uncoded BPSK and ML detection,
as in the amplitude detection case. For the phase detection MIMO system, we use two
modulation schemes: BPSK without spatial multiplexing (rate R 1 bit/s/Hz) and
Ex-BPSK with spatial multiplexing (rate R NT bits/s/Hz). The first modulation
scheme consist of the symbols r1, . . . 1sT and
r1, . . . , 1sT and is equivalent to using
only one transmit antenna and the symbols t1u. The channel has i.i.d. Gaussian
entries with unit variance and we use the SNR per bit Eb {N0 .
Diversity of D1i and D2i : First we provide simulation results for the diversity order
of D1i and D2i in Fig. 4.9a. Recall that it was not possible to compute the diversity
order of D2i analytically. As shown in Fig. 4.9a, similar to amplitude detection, D1i has
diversity order equal to 1 and D2i has diversity order equal to 21 .
22 MIMO: Fig. 4.9b depicts the average BER of a 2 2 system. Phase detection
exhibits diversity equal to NR 2 when R 1 bit/s/Hz, and diversity equal to N2R
when R 2 bits/s/Hz. The BER of linear detection and phase detection with R 1
bit/s/Hz are identical, however the linear detection system conveys twice the bits. In
6
The diversity order of NR 4 is measurable only for larger Eb {N0 . However, in the regime depicted
in Figs. 4.8a, 4.8b, the respective curves have the same slope and we conjecture that amplitude
detection with R 1 bit/s/Hz has also diversity 4.
56
4.4 Performance Results
Linear BPSK 2x3, R=2 bits/sym Linear BPSK 3x3, R=3 bits/sym
Phase Det. 2x3, R=2 bits/sym Phase Det. 3x3, R=1 bits/sym
−1 −1
10 Phase Det. 2x3, R=1 bit/sym 10 Phase Det. 3x3, R=2 bit/sym
Linear BPSK 2x4, R=2 bits/sym Phase Det. 3x3, R=3 bit/sym
Phase Det. 2x4, R=2 bits/sym Linear BPSK 3x4, R=3 bits/sym
Phase Det. 2x4, R=1 bit/sym Phase Det. 3x4, R=1 bits/sym
−2 −2 Phase Det. 3x4, R=2 bit/sym
10 10 Phase Det. 3x4, R=3 bit/sym
Average BER
Average BER
−3 −3
10 10
−4 −4
10 10
−5
10
−10 −5 0 5 10 15 20 25 −10 −5 0 5 10 15 20 25
Eb/N0 Eb/N0
(a) (b)
Figure 4.10: (a) BER plot for a 2 3 and 2 4 MIMO system, (a) BER plot for a 3 3
and 3 4 MIMO system.
2NR MIMO: Fig. 4.10a depicts the average BER of a 2 3 and 2 4 system,
with and without spatial multiplexing. As predicted by the analysis of the PEPs, the
diversity order with spatial multiplexing is only N2R 1.5 and N2R 2, respectively.
Without spatial multiplexing, the full diversity of NR 3 and NR 4, respectively, is
achieved.
3NR MIMO: Fig. 4.10b finally depicts the same comparison for a 3 3 and 3 4
system. This time, we have three levels of spatial multiplexing, yielding the uncoded
rates R 1, 2, 3 bits/s/Hz. The observation for the R 1 bits/s/Hz and R 3
bits/s/Hz cases are as expected—the diversity order equals NR and N2R , respectively.
The
! amodulation
) for the R 2 bits/s/Hz case has the form s ra11 , a11 , a12 sT , a1i P
Es{NT , such that two independent data streams are transmitted. The resulting
NR
diversity order is again 2
.
57
Chapter 4 Modulation and Maximum Likelihood Detection
We analyzed the performance of the ML detector of MIMO amplitude and phase de-
tectors. Contrary to linear MIMO systems, the nonlinear MIMO systems loose half
the available receive diversity when a spatial multiplexing modulation alphabet is em-
ployed. This result does not imply that multiple transmit antennas are useless for such
systems, but it rather shows that a performance penalty is induced by the nonlinearity
of the receiver. The mechanism that leads to the diversity is not identical for amplitude
and phase detection, although there are certain similarities. The MIMO amplitude de-
tector losses the full diversity the moment the uncoded rate is higher than 1 bit/s/Hz,
even in a SISO system. The MIMO phase detector on the other side can achieve the full
diversity by using M-PSK in the SISO/SIMO case, i.e., uncoded rates higher than 1
bit/s/Hz are possible without sacrificing the diversity. The diversity penalty for MIMO
phase detection occurs the moment more than one transmit antennas are used.
58
Chapter 5
Channel Estimation
In this chapter we make a further step towards practical nonlinear MIMO systems. In
Chapter 4 we assumed the receiver has perfect knowledge of the channel. However, it
is not clear how a MIMO amplitude or phase detector can estimate a complex-valued
channel, or whether estimation of the full channel is necessary at all. In this chapter
we investigate channel estimation techniques for MIMO amplitude and phase receivers
in Sections 5.1 and 5.2, respectively, but focus mainly on the former. We present
three novel estimation schemes for each receiver type and compare them through the
performance of the ML detector that makes use of the estimated channel. In Chapter 6
we will actually implement the proposed schemes for amplitude detection in a real
system. Section 5.1 has been published in [22] and [25].
So far, we either assumed perfect CSI is available at the receiver (cf. Section 3.1 and
Chapter 4), or used a linear minimum mean square error (MMSE) estimation model to
include noisy CSI (cf. Section 3.2). Meanwhile, it has not been clear how a MIMO am-
plitude detector can estimate the channel coefficients. In fact, estimating the phases of
the channel coefficients is indeed not possible, since the phase information is removed
when taking the amplitude of the received signal. However, as we will show in Sec-
tion 5.1.1, complete knowledge of the channel coefficients is not required. The MIMO
amplitude ML detector can operate using only knowledge of the norm of the channel
coefficients, and the norm of linear combinations of the channel coefficients. We refer
to this knowledge as sufficient CSI, and propose channel estimation schemes that allow
the MIMO amplitude detector to estimate the necessary CSI without relying on other
59
Chapter 5 Channel Estimation
means, like obtaining CSI on a feedback channel. Our proposed schemes decompose
the estimation problem in several scalar estimation problems. We analyze different
estimation techniques for the scalar estimations in Section 5.1.2, such as maximum
likelihood, Bayesian and method of moments estimation, following a similar approach
to [43]. Finally, we propose three estimations schemes in Section 5.1.3.
We start by examining the ML detector that uses perfect CSI, given in (4.4) and (4.29):
¹
NR
spML argmax f py|zq argmax f pyi |zi q . (5.1)
s PS s PS
i 1
2 2
2yi yi σ2xi,s
2
σw
e w I0
2yi xi,s
σw2
, xi,s eTi Hs . (5.3)
Here, xi,s |zi| is the noiseless received signal at the ith antenna. The index s makes
dependency on the specific s explicit.
Observing (5.1) and (5.3), it is apparent that the ML detector does not require
complete knowledge of all channel coefficients hij . Actually, it is sufficient to have
knowledge of all elements in
(
X xi,s eTi Hs i 1, . . . , NR, s P S . (5.4)
60
5.1 Channel Estimation for Amplitude Detection
consider Ex-OOK. In this case, for those elements of S that have a single non-zero entry
at the kth position, i.e. sk r0, . . . , 0, α, 0, . . . , 0sT , the receiver requires knowledge of
that is, knowledge of |hij |, i, j. For the remaining symbols sn P S ztsk u that have two
or more non-zero entries, the receiver requires knowledge of:
¸NT
xi,sn
hil sn
r s
l ,
i P t1, . . . , NRu , sn P S ztsk u (5.6)
l 1
The cardinality of the set of the sufficient CSI is |X |
p|S | 1q NR 2N 1 NR T
and depends on the modulation alphabet1 S. In the case of Ex-OOK, |X | grows expo-
nentially with NT , or in other words, grows exponentially with the uncoded rate NT . If
for instance we use a rate 1 bit/s/Hz alphabet,
comprising
the symbols S 1 t0, α 1u,
°
l1 hil i P t1, . . . , NR u, that is, |X | NR .
the required channel knowledge is only α N T
The exponential relation is due to the exponential complexity of the ML detector: since
for every possible symbol s P S we have to compute the conditional probability, and
the number of symbols in S is exponential in NT , the amount of channel knowledge
for each assumption also depends exponentially on NT . Note that we could construct
all entries in X if we knew all NT NR complex-valued channel coefficients.
We consider again Ex-OOK and gather the sufficient CSI quantities in a vector θ.
We neglect α, since it is known to the receiver:
The structure of θ continues as shown above, including progressively all linear combi-
nations of hij across the rows of H. Note that the θi ’s are not independent of each
other. Consider for example the parameters
1
The sufficient CSI does not include the trivial all-zero symbol s 0.
61
Chapter 5 Channel Estimation
These three parameters are connected through the triangular inequality as follows:
Similar inequalities relate the rest of the parameters in θ. The dependencies in θ sug-
gest that joint estimation of θ is necessary for optimal estimation, albeit complicated.
If we considered ML estimation, the joint estimation problem would read as
The following description applies to estimation of any quantity of the sufficient CSI.
Without loss of generality, assume the symbol s λ r1, . . . , 1, 0, . . . , 0sT , λ P R ,
with L non-zero entries is transmitted. The received symbol at the ith antenna can be
written as
yi |hλ wi | ||h|λ wi1 | |θλ wi1 |, (5.13)
°
where h Li1 hi and λ is a known scalar. Note that only the number L will be of
interest, and not the exact position of the non-zero elements in s. We wish to estimate
the parameter θ |h| ¥ 0. Collecting all N received symbols at the ith antenna, we
obtain
y |θλ w1 | P RN 1 , (5.14)
where λ λ1, w1 CN p0, σw2 IN q and we omit the index i for the sake of brevity.
The noise added at different time instances is statistically independent, which implies
that the receiver possesses N independent observations of the unknown parameter.
62
5.1 Channel Estimation for Amplitude Detection
Note that y in (5.14) refers to several instances of the received signal at the ith an-
tenna, contrary to the previous section, where y described the received vector at all
receive antennas at a given time instant. Assuming θ is a non-random parameter, the
distribution of y is given by:
¹
N ¹
N
2yi yi σλ2 θ
2 2 2
f py; θq f pyi ; θq
2λyi θ
2
e w I0 . (5.15)
i 1 σw
i 1
σw2
Bθ i1 I0 2λθy
σ
i
2
i
(5.16)
w
that maximizes the likelihood function (there might be several roots). Eq. (5.16) is
nonlinear with no closed form solution. We will numerically compute the MLE and
compare it with our complexity reduced estimators in Section 5.1.4
Cramér-Rao Lower Bound : The lower bound of the variance for estimating θ using
N independent observations of the receiver output y, is given by
CRLBpθq
1
N ipθq
(5.17)
63
Chapter 5 Channel Estimation
where ipθq is the Fischer information for one observation (sample), given by
ipθq E
B ln ppy; θq
2
Bθ (5.18)
» 8 8λ2 y 3 I2 2λθy
4 2
y2 σλw22 θ2
4λσ4θ
1 2
σw
e dy. (5.19)
2λθy
w 0 σw6 I0 2
σw
The Bayesian approach takes the distribution of θ into account. In our scenario,
h CN p0, Lσh2 q and |h| is Rayleigh distributed, like
2θ θ2 {L
f pθq e , θ ¥0 (5.20)
L
where we set σh2 1. The Bayesian estimate can be computed in closed form only
for N 1 sample. The conditional distribution f py |θq is identical to (5.15). The
distribution of y is Rayleigh, computed as
»8
Lλ22y σ2 e
y2
f py q f py |θq f pθqdθ 2
Lλ2 σw
. (5.21)
0 w
The posterior distribution of θ follows from ppθq, (5.15), (5.21) and Bayes rule
θ θ2 2ν 2
f pθ|y q θν
e 2σ I0 (5.22)
σ2 σ2
64
5.1 Channel Estimation for Amplitude Detection
ν2
θpMMSE E rθ|ys πσ 2
1 F1 1{2; 1; 2σ
? 2 2
Lλ2 y 2
a 2σw Lπ
1 F1 1{2; 1; σ2 pLλ 2 σw2 q
, (5.24)
2 Lλ σw2 w
using the mean of a Rice distribution. 1 F1 pq is the confluent hypergeometric function.
The conditional mean is an ascending
function of the observation y. The estimator
yields values in the range θpmin , 8 , where
?
θpmin a
σw Lπ
, (5.25)
2 Lλ2 σw2
obtained by setting y 0 in (5.24). At high SNR, the minimum value tends to zero
2
σw
lim θpmin
Ñ0
0, (5.26)
and the Bayesian estimator can estimate all possible values of θ in the range r0, 8s.
At low SNR however, the minimum value is
?
2
σw
lim θpmin
Ñ 8
Lπ
2
0, (5.27)
which equals the mean of the a priori parameter distribution, i.e., the mean of the
Rayleigh distribution in (5.20). In this case, the estimator relies only on the a priori
?
knowledge of the parameter, and ignores the noisy observation. Values lower than
Lπ {2—which are less likely—can not be estimated. This lower bound for low SNR
will prove useful later. For N ¡ 1, no closed form solution exists. We choose to
average the individual Bayesian estimates, although this is not optimum as we will see
in Section 5.1.4.
A popular estimation method which often leads to simple and consistent estimators
uses the moments of the given distribution. In our case, we will consider the first two
65
Chapter 5 Channel Estimation
µ1 E ry s
2
w
1 F1 12 ; 1; λσ2θ , (5.28)
w
µ2 E y2 σw2 λ2 θ 2 . (5.29)
where we assume the inverse fn1 pq exists. In case of θpµ1 , this inverse must be computed
numerically. However, the second moment yields a very simple estimator:
a
θpµ2 g2pµp2q λ1 p2 σw
µ 2, p2 σw
µ 2
¥ 0. (5.31)
It can be shown that θpµ2 is a consistent estimator. Using Jensen’s inequality, we obtain
1a
E θpµ2 E
λ
p2 σw
µ 2 (5.32)
1a
¤ λ
E rµ
p2 s σw
2 θ (5.33)
meaning that θpµ2 is biased. However, using a first-order Taylor expansion about E rµ
p2 s
µ2 for high N, we get [45]
dg2 pµ
p2 q
θpµ 2 g2pµ2q dµ
pµp2 µ2q
p2 µp2 µ2
(5.34)
66
5.1 Channel Estimation for Amplitude Detection
which means that the variance becomes asymptotically zero when σw2 Ñ 0, and thus
θpµ2 is consistent.
p2 σw
When µ 2
0 the estimator can not be computed. This situation can occur at
low SNR, in which case we can set the estimate for θ to zero. This means that there
is no propagation channel. Although this prohibits communication, the effect on the
BER is negligible. This is because this effect occurs at low SNR, where performance is
anyway very poor due to large noise. In this regime, knowing the channel well, poorly
or not knowing the channel at all makes no real difference. Another possibility to
avoid using a detrimental channel h 0 is to resort to the Bayesian estimate whenever
p2 σw
µ 2
0, since, as we saw, the Bayesian estimate will deliver a non-zero lower
bounded value at low SNR.
67
Chapter 5 Channel Estimation
SEES
Ntot |S 1|N p2N 1qN.
T
(5.37)
During the channel estimation phase, the receiver performs NR p2NT 1q scalar esti-
mations. Note that the exclusion of the all-zero word from the pilot implies that we
transmit slightly more power per symbol in the pilot phase. To compensate for this,
a a
we use the value β p2NT 1qEs {pNT 2NT 1 q instead of α Es {NT for the pilot
symbols (for Ex-OOK), such that the average energy per pilot symbol remains Es .
SEES is very simple to implement and leads to very good performance, as we will
SEES
se in Section 5.1.4. The major disadvantage of SEES is the fact that Ntot scales
exponentially with NT , hence with the uncoded rate when we use spatial multiplexing
(Ex-OOK). Thus, use of the SEES for large MIMO systems quickly becomes pro-
hibitive.
One way to increase the uncoded rate without changing the number of transmit
antennas and the pilot length is using M-OOK. In this case, it suffices to transmit
the first 2NT 1 symbols. The additional terms in X can be computed using the
already estimated terms and the cosine rule, while the rate increases to NT log2 M.
For example, if we use three modulation levels t0, ǫ, ζ u, ǫ, ζ P R, we can compute
combinations of the form |ǫh1 κh2 | as follows
a
|ǫh1 κh2 | ǫ2 |h1 |2 κ2 |h2 |2 2ǫκ|h1 ||h2| cos =ph1 , h2 q,
where the angle in =ph1 , h2q is computed from |h1|, |h2|, |h1 h2 | and the cosine rule.
68
5.1 Channel Estimation for Amplitude Detection
The same holds when using more modulation levels. In essence, we use the property
=ph1 , h2q =pǫh1 , κh2q. In the next to sections we present two schemes that try to
control the exponential growth of Ntot using different techniques.
The GAES scheme takes advantage of the dependency among entries of the sufficient
CSI, which stem from the geometry of the channel coefficients on the complex plane.
The GAES pilot employs all symbols sGAES P S A Y S B:
SA tγe1 , . . . , γeN u,
T
(5.38)
SB ts | s γ pei ej q for any i j u. (5.39)
The set S A consists of all NT symbols with a single non-zero entry, and S B consists
a
of all N2T combinations of two non-zero entries. We choose γ pNT 1qEs {p2NT q
such that the average energy per pilot symbol equals Es . GAES estimates only the
absolute value of the channel coefficients |hij |, i, j using the pilots in S A , and the
absolute value of all combinations of two channel coefficients, using the pilots in S B .
This corresponds only to a subset of the sufficient CSI X . The remaining entries, which
are the absolute values of sums of three or more coefficients, are then computed using
the already obtained estimates and some geometric properties. This modification to
SEES leads to a pilot of total length
A B
GAES
Ntot S S N NT
NT
2
N 12 NTpNT 1qN, (5.40)
GAES concept The GAES scheme is based on the following observation. Consider
three complex-valued scalars h1 , h2 , h3 . Suppose we know the following quantities:
Note that these quantities can be estimated directly from the pilot of GAES. Using
these quantities, we can compute the angle φ1,2 between h1 and h2 , and φ2,3 between
69
Chapter 5 Channel Estimation
h13
h1
φ2,3
h3
φ2,3
h2
φ1,2 h2
h1 symmetry
φ1,2
h2
h2 φ2,3
h3
h1 φ2,3
h13
Figure 5.1: Sketch of all possible relative positions of h1 , h2 and h3 on the complex plane.
Now, consider the vectors h1 , h2 , h3 positioned on the complex plane according to the
known lengths of the vectors and the computed angles, as depicted in Fig. 5.1. The
angles computed by the cosine rule are positive, but the sign does not give any in-
formation regarding the direction of the angle on the complex plane. Without loss of
generality, assume h1 is fixed and acts as our reference. Using φ1,2 and |h2|, we draw
two possible vectors h2 . Similarly, using φ2,3 and |h3 | we draw two possible vectors h3
with respect to h2 , i.e., we obtain four possible vectors2 h3 . In total, the ambiguity of
the angles φ1,2 and φ2,3 leads to four possible relative positions between h1 , h2 and h3 .
As shown in Fig. 5.1, the topologies are symmetric around h1 . We neglect the lower
part and only consider the two possible topologies on the upper half of Fig. 5.1. Due
to the phase ambiguities, there are two possible lengths for |h1 h2 h3 |, namely |h1 |
and |h2 |. However, if we also know |h1 h3 | — which can also be computed directly
from the pilot of GAES — we can resolve this ambiguity. Computing the angle φ1,3
between h1 and h3 we can determine whether h3 or h13 is the correct vector. Thus, we
can uniquely determine the relative topology between h1 , h2 and h3 , and consequently
2
We name two of them h13 in order to distinguish them.
70
5.1 Channel Estimation for Amplitude Detection
compute |h1 h2 h3 |. Summarizing, the absolute value of the sum |h1 h2 h3 | can be
uniquely calculated using the individual absolute values |h1|, |h2|, |h3| and the absolute
values of the sums: |h1 h2 |, |h2 h3 | and |h1 h3 |.
The above observation suggests that we can compute all entries of the sufficient CSI
using only the estimates from S A and S B . We devise an algorithm which computes
|h1 h2 h3 | for all possible topologies, and uses |h1|, |h2|, |h3 |, |h1 h2 |, |h2 h3 |
and |h1 h3 | as input arguments. We call this the component block of GAES (cGAES)
algorithm. The absolute value of sums of more than three channel coefficients can be
computed iteratively, by splitting the computation in smaller steps, until the cGAES
algorithm can be applied. E.g., |h1 . . . h4 | can be computed as follows: first ßset
h1 h1 h2 . Then, cGAES can be called with arguments:
$
'
'
'
'
| | |h1 h2 |
h1
'
'
'
'
'
'
| |
h3
'
'
| 4|
&h
. (5.43)
'
'
'
'
|h3 h4 |
'
'
'
'
'
'
|h1 h3 | |h1 h2 h3 |
'
'
%|h1 h | |h h2 h4 |
4 1
The last two absolute values are sums of three coefficients, and can be readily computed
by two additional calls of cGAES. Using the algorithm iteratively, one can compute
any element of the sufficient CSI X .
We describe the algorithm cGAES in the following section, while pseudocode is
given in Algorithm 1. The algorithm is first designed for perfect estimates, and then
modified to account for noisy estimates. In the noiseless case, cGAES determines the
relative position of the vectors h1 , h2 and h3 , albeit without knowing the exact phases.
This knowledge can not be extracted due to the amplitude detection operation at the
receiver. In realistic systems however, cGAES takes noisy estimates as input values.
h1 |, |y
Automatically, the geometrical properties are no more valid, i.e., |y h2| and |h{
1 h2 |
might not even form a triangle, thus violating the triangular inequality. Hence, the
algorithm operates in a probabilistic manner. Further measures are taken in order to
account for noisy estimates, as described in Appendix D.2. A drawback of GAES is
the error propagation as a consequence of calling cGAES iteratively. As we will see in
Section 5.1.4.2, the estimation quality of sums of many channel coefficients degrades
as the number of summands increases.
71
Chapter 5 Channel Estimation
F
2.5
h3
2 G
B
h +h
h3 H 2 3
1.5
φ
P =ODG
AD
12,3 yes
no
φ D
2,3 δ h +h +h h2
P =ADG OD
P =GDH
1 2 3
1
θ OD
C
no yes no yes
h1+h2 h3
0.5
h2
P =ADO HD
DG
P =ODG
no yes no yes
E h1+h3 δ
0
O A1.1 A1.2 A2.1 A2.2 B1 B2
h3
φ h1
−0.5 1,3
φ1,2
A
−1
−2.5 −2 −1.5 −1 −0.5 0 0.5 1
(a) (b)
Figure 5.2: (a) Exemplary topology for case A1.1 of cGAES (h1 1.8646 j0.7706,
h2 0.7004 j1.8945, h3 0.1578 j0.7665), (b) Logical diagram for cGAES
algorithm.
The choice of φ12,3 is arbitrary—we could for example combine h1 h3 in (5.44) and
use φ13,2 . All quantities in (5.44) are known, except for φ12,3 which we have to compute.
First, some auxiliary quantities are required. Let φi,j be the angle between hi and hj
72
5.1 Channel Estimation for Amplitude Detection
as computed by the cosine rule in (5.42). We compute φ1,2 , φ1,3 and φ2,3 in line two
of Algorithm 1. Next, the auxiliary angles θ and δ are also computed in lines 3 and 4.
Note that φ1,2 , θ and δ are the three angles of the triangle △OAD. The calculation
of φ12,3 depends on the relative position of h1 , h2 and h3 on the complex plane. We
identify six different topologies and compute Φ12,3 piq, i 1, . . . , 6 for each one of them
(lines 5 to 16). The algorithm then requires a means to distinguish which topology
is the correct one. This is achieved by recalculating the angle between h1 and h3 ,
Φp1,3 piq, in such a way that Φ12,3 piq is involved. The new angles Φ
p1,3 piq are computed
p1,3 piq is compared
in lines 6, 8, 10, 12, 14 and 16 for each possible topology. Finally, Φ
with the correct φ1,3 (line 22), and the actual topology is identified. In line 23, the
correct angle φ12,3 is chosen, and computation of |h1 h2 h3 | follows in line 24 using
(5.44). The algorithm is deterministic when the input arguments are noiseless, i.e.,
perfect estimates. For noisy input arguments, the comparison in line 22 yields the
most probable topology.
73
Chapter 5 Channel Estimation
topologies will be considered. Similarly, Φ12,3 p2q and Φ12,3 p4q are opposite, and only
one topology yields a positive angle.
The algorithm described above operates on noiseless estimates for the input ar-
guments. Further measures are required when noisy estimates are used (see Ap-
pendix D.2). Notice that the algorithm can not handle zero estimates for any of
the arguments, since then the various triangles are not defined. In order to avoid this,
we use a modified MoM estimator, that resorts to the Bayesian estimate whenever
p2 σw
µ 2
0 (see Section 5.1.2.3).
74
5.1 Channel Estimation for Amplitude Detection
Estimation of the exact phases of the channel coefficients is not possible due to the
absolute value operation of the amplitude detector. However, it is possible to estimate
the phase difference between different channel coefficients. ARES makes use of this
property in order to estimate a channel matrix. We explain the principle of ARES
using the following example. Consider the computation of some x P X :
¸K ¸ ¸ ¸
| | ejφir hik | | ejφir hik | |
jφir
hik e hir hir hir hrik ,(5.45)
k 1
k r
k r
k r
where K takes any value K P t2, . . . , NR u and φir =hir . The absolute value (5.45)
can be written as the sum of an arbitrary |hir |, which is the absolute value of a chosen
reference channel coefficient, and the remaining channel coefficients rotated appro-
priately. The phases of the new coefficients4 hrik ejφir hik , are actually the phase
differences between the coefficients and the reference hir :
The amplitude detector can actually estimate these phase differences, and along with
estimating |hrir |, the complex-valued hrir are estimated. Computation of the sufficient
CSI follows subsequently as in (5.45). Summarizing, ARES chooses some reference
transmit antenna, say the rth antenna, first estimates all |hir | P R , then estimates all
hrik P C, i, k r and finally constructs an estimated channel matrix H xARES . At the
end, the real channel differs from the estimated one by element-wise multiplication with
a matrix Ψ , comprising the phases of the reference channels on each receive antenna
(in the noiseless case):
H H
xARES d Ψ H
xARES d rejφ 1r
, . . . , ejφNR r sT b 1T
NT . (5.47)
The presence of the phase shifts in Ψ has no impact on the computation of the sufficient
CSI X , as shown in (5.45).
ARES Pilot: The pilot sequence in ARES is composed of all symbols sARES P
4
The superscript ‘r’ denotes the reference to the rth transmit antenna.
75
Chapter 5 Channel Estimation
S A Y S B Y S C defined as:
SA tδe1 , . . . , δeN u, T
(5.48)
SB ts | s δper ek q for any k ru, (5.49)
SC ts | s δper jek q for any k ru, (5.50)
a
where δ p3NT 2qEs{p5NT 4q is chosen such that every pilot symbol has average
energy Es . Like in GAES, all NT symbols with a single non-zero entry are transmitted
in S A . This enables estimation of |hij |, i, j. Next, one reference transmit antenna, say
r P t1, . . . , NT u, is chosen. Then, all NT 1 symbols in S B which have two non-zero
entries are transmitted: one fixed at the rth position, and one at the kth position,
k r. This allows estimation of all combinations: |hir hik |, k r. Finally, the
previous NT 1 symbols are transmitted again in S C , but using a π {2 phase shift on
the secondary antenna, i.e., a j is transmitted on the kth position. This final step
enables estimation of the phase of all channel coefficients with respect to the reference
channel coefficient. At the end, the length of the pilot is
ARES
Ntot pNT 2pNT 1qq N p3NT 2qN, (5.51)
ARES
if we assume that N repetitions are used for every pilot symbol. Note that Ntot
grows only linear with NT .
ARES Estimation: After receiving S A , the receiver can estimate SijA |y hij |,
j 1, . . . , NT. Let the channel coefficient from the reference antenna be hir ziejφir ,
zi P R . So far, we have an estimate for Sir
A
zpi |y
hir |. Now, we make the following
definitions
hrik ejφ ir
hik ℜtejφ ir
hik u jℑtejφr hik u Rik jIik , (5.52)
for all k r. Hence, we first rotate hik by φir , and then compute the real and
imaginary parts. The latter are now computed in a new coordinate system, where hir
lies solely on the new real axis, as depicted in Fig. 5.3. Furthermore, notice that
jφ b
|hik | | |
hrik e r hik 2
Rik 2
Iik . (5.53)
In the following we will estimate Rik and Iik , and hence compute all channel coeffi-
cients in the ith receive antenna in complex-valued form, with respect to the reference
76
5.1 Channel Estimation for Amplitude Detection
ℑtu
hik
Iik hir
ℜtu
Rik
φir
Figure 5.3: Example: let hir be the reference channel coefficient. hik in Rik and Iik is
decomposed with respect to hir .
Rik ik
pS B q2 pSirAq2 pSikA q2 . (5.58)
2S A ir
Eq. (5.58) describes the computation of Rik using the estimates from S A and S B . The
third part of the pilot is additionally required for computing Iik . Following the same
reasoning, using S C we estimate
jφ
|hir jhik | e r zi p jejφr hik q |zi jRik Iik | (5.59)
b
pzi Iik q2 Rik2
b
zi2 2zi Iik |hik |2 , (5.60)
for all k C
r. Let the estimate be Sik |hir{jhik |. Assuming again error-free
77
Chapter 5 Channel Estimation
estimates, we obtain
Iik ir
pS Aq2 pSikA q2 pSikC q2 . (5.62)
2S A ir
This concludes the computation of hrik . In the noisy case, we compute estimates
p
hr
ik Rpikj Ipik using (5.58), (5.62) and the noisy estimates Sik
A,B,C
. However, no-
tice that we had already computed directly the estimate for |hik | using S A . As a last
y
step, we normalize p hrik as follows:
p
hrik SikA ej=pRp
ik j Ipik q. (5.63)
Hence, we only use the phase that results from (5.58) and (5.62). Finally, combining
all estimates p
hrik and Sir
A
we can construct the matrix estimate
.. .. .. .. ..
. . . . .
xARES
H phi1 . . . phipr1q Sir phipr 1q . . . phiNT
r r A r r .
(5.64)
.. .. .. ... ...
. . .
The entries of the rth column are real-valued, since we used the rth transmit antenna
as our reference, and cancelled the phases of all hir . The rest of the entries are complex-
valued, and correspond to coordinate systems which are rotated by φir on each receive
xARES is not an estimate of the channel matrix H, it
antenna, respectively. Although H
can be used to compute all entries in X , as shown in (5.45).
78
5.1 Channel Estimation for Amplitude Detection
1
10
CR lower bound ML Estimation
2
10 ML Estimation MoM Estimation
MoM Estimation 0 Bayesian Estimation
10
N = 1, 2, 4, 10, 20, 50
N = 1, 2, 4, 10, 20, 50 −1
0
10 10
Variance
MMSE
−2
10
−2
10
−3
10
−4
−4 10
10
−5
10
−10 −5 0 5 10 15 20 25 30 −10 −5 0 5 10 15 20 25 30
SNR SNR
(a) (b)
Figure 5.4: (a) Variance of MLE and MoM Estimator. The CRLB is valid above SNR 15
dB, (b) MMSE comparison of MLE, MoM and Bayesian estimators.
In this section we assess the performance of the various scalar estimation methods
presented in Section 5.1. Subsequently, we use our proposed schemes in a system
simulation and evaluate the performance of the ML detector that uses an estimated
channel in terms of BER.
Fig. 5.4a depicts the variance of the MLE and MoM estimator together with the CRLB.
The SNR is defined as SNR σE2s . The two estimators have practically the same
w
variance, which means that although θpµ2 is a very simple estimator, it is as powerful
as the MLE. The CRLB is violated at low SNR, since the two estimators are biased.
However, at high SNR and for high number of samples the bound is tight.
Fig. 5.4b depicts the MMSE of the MLE, the MoM and the Bayesian (L 1)
estimators for different block lengths. Like in Fig. 5.4a, the performance of MLE
and MoM is practically identical. For N 1, the Bayesian estimator yields the best
estimate for all SNR, since it is MMSE optimal. For N ¡ 1, we average the individual
Bayesian estimates, although this is not optimum. As a result, the MoM and ML
estimates progressively improve and outperform the Bayesian estimate. At high SNR
the estimators perform similarly. The Bayesian estimator performs very good with
79
Chapter 5 Channel Estimation
3
10
Asymptotic CRLB
Asymptotic variance of MoM Estimator
2
10
1
10
Asymptotic Variance
σ2w=[1, 0.1, 0.01, 0.001]
0
10
−1
10
−2
10
−3
10
0 0.5 1 1.5
θ
respect to the MMSE for small N and at low SNR. In the next subsection, we will see
how this gain reflects to the BER performance.
Fig. 5.5 depicts the asymptotic variance N varpθpµ2 q of θpµ2 computed in (5.36), and
N CRLBpθq ip1θq computed in (5.19), as a function of the estimation parameter, for
different SNR. The MoM Estimate achieves asymptotically the CRLB. The estimation
variance grows unbounded when θ goes to zero, since in this case the receiver observes
solely noise, and saturates at σw2 {2 as θ grows.
22 MIMO system: First we consider a 2 2 MIMO system and SEES. GAES and
ARES are not applicable for NT 3. Fig. 5.6a depicts the BER performance with
perfect CSI and with estimated CSI using SEES. We compare the behavior of a non-
linear MIMO system that performs channel estimation with a linear MIMO reference
system that uses BPSK and spatial multiplexing, yielding the same uncoded rate of
R 2 bits/s/Hz. Note that the nonlinear MIMO system is not an alternative to a
linear MIMO system, since its performance is clearly inferior. The linear MIMO system
performs MMSE channel estimation using a properly scaled submatrix P P CNT Ntot of
the DFT matrix as the training sequence. The estimated channel is given by (cf. [46])
1
xLMMSE
H NRY NR P H P σw2 I P H. (5.65)
For a fair comparison, we have allocated the same resources for the pilot in
80
5.1 Channel Estimation for Amplitude Detection
1.5
−2
10
1
Linear MIMO, perfect CSI
Linear MIMO, LMMSE
Nonlin. MIMO, pefect CSI 0.5
Nonlin.MIMO, MoM
−3
Nonlin. MIMO, MMSE
10 0
−10 −5 0 5 10 15 20 25 3 6 9 12 15 18 21 24 27
Eb/N0 Ntot
(a) (b)
Figure 5.6: (a) BER of a 2 2 MIMO system. Ntot 3, (b) MIMO 2 2 system at
BER=102 . SNR penalty of various estimation methods with respect to perfect
CSI, as a function of Ntot .
both systems. Recall that for N 1, the nonlinear MIMO system requires
ARES
Ntot p2NT 1qN 3 training symbols. In Fig. 5.6a we use Ntot 3 for both
systems. This translates to N 1 for the nonlinear MIMO system. The SNR per bit
is given by Eb {N0 SNR{R. We see that both systems exhibit a similar SNR penalty
when the channel is estimated. The Bayesian estimator is insignificantly better than
the MoM estimator at low SNR. Hence, the superiority of the Bayesian estimator at low
SNR, as seen in Fig. 5.4b, is lost in the BER performance. This indicates that at low
SNR the performance degrades mainly due to large noise, and the channel estimation
quality has a much smaller impact.
In Fig. 5.6a we used the smallest value for Ntot . As Ntot increases, the performance
gradually improves. This is captured in Fig. 5.6b, where the SNR penalty with respect
to perfect CSI is plotted as a function of Ntot , at a BER of 102 . The performance
gap of the nonlinear MIMO system with channel estimation is bigger than that of a
linear MIMO system. Note that the linear MIMO system benefits from the proper
design of the pilot sequence, which exhibits good correlation properties. This leads
to an additional SNR gain, as seen in Fig. 5.6b. Unfortunately, using such trainings
sequences with the nonlinear MIMO system is not possible. Furthermore, we observe
that the Bayesian estimator offers practically no advantage at small values of Ntot , and
slowly performs worse than the MoM estimator as the length of the pilot increases.
Concluding the discussion on the scalar estimators, we can say that the simple MoM
81
Chapter 5 Channel Estimation
Table 5.1: Simulation parameters for SEES, GAES, ARES, and NT t3, 4, 5u.
NT pNSEES, NGAES, NARES q Ntot
SEES GAES
Ntot ARES
Ntot pζ, ξq
(1,1,1) 7 6 7 (7/6, 1)
3 (6,7,6) 42 42 42 (1, 1)
(10,12,10) 70 72 70 (70/72, 1)
(1,1,1) 15 10 10 (3/2, 3/2)
4 (2,3,3) 30 30 30 (1, 1)
(4,6,6) 60 60 60 (1, 1)
(1,1,1) 31 15 13 (31/15, 31/13)
5 (1,2,2) 31 30 26 (31/30, 31/26)
(2,4,5) 62 60 65 (62/60, 62/65)
scalar estimator is the preferable choice for the complexity-reduced, low-power MIMO
amplitude detector. The resulting BER performance is within 1 dB from the ML
detector with perfect CSI, for moderate pilot lengths.
Larger MIMO systems Now we compare the performance of SEES, GAES and
ARES for different MIMO scenarios with NT ¥ 3. In order to achieve a fair com-
parison, we try to use the same pilot length for all schemes, such that the spectral
efficiency is the same. When this is not possible, we normalize the transmit power
of the pilot symbols such that all schemes use the same transmit energy during the
estimation phase. For SEES we use the MoM scheme for the scalar estimations, since
this scheme carries the most advantages. For GAES and ARES we use a mixed mode:
the MoM estimate is first computed. If the estimate is zero, the Bayesian estimate
is computed and used instead. This way, we avoid using zero channels and at the
same time, we benefit from improved performance for high values of N (cf. 5.4b). For
ARES we always choose the first transmit antenna as the reference antenna, r 1.
For equally fading channels, this fixed choice has no impact on performance.
Using the same pilot length for all schemes is not always possible. E.g, for NT 3,
we obtain: pNtot
SEES GAES
, Ntot ARES
, Ntot q p7NSEES , 6NGAES, 7NARESq. Equalizing the pilot
length for all schemes suggests choosing common multiples of 6 and 7. The simplest
mode, NSEES NGAES NARES 1, penalizes GAES which uses one pilot symbol
less than SEES and ARES. In this and similar cases henceforth, we will scale the pilot
symbols such that all schemes transmit totally the same energy in the pilot phase,
SEES
hence yielding a fairer comparison. In the above case, Ntot ζNtot
GAES
ξNtot
ARES
? ,
where ξ 1 and ζ 7{6. The pilots of GAES and ARES will be multiplied with ζ
82
5.1 Channel Estimation for Amplitude Detection
−1
10
BER
−2
10
SEES (1,1,1)
GAES (1,1,1)
ARES (1,1,1)
SEES (6,7,6)
GAES (6,7,6)
ARES (6,7,6)
−3 SEES (10,12,10)
10
GAES (10,12,10)
ARES (10,12,10)
−5 0 5 10 15 20 25
Average SNR (dB)
?
and ξ, respectively, and hence all three schemes will transmit the same amount of
energy during the training phase. Notice that we implicitly used the fact that each
pilot symbol was already normalized to use Es energy on average.
33 MIMO: First we start with the smallest MIMO system where GAES and
ARES are applicable. For the 3 3 MIMO system, we choose the parameters found
in Table 5.1 for NT 3. In the following plots, dotted, dashed and solid curves
stand for SEES, GAES and ARES, respectively. Curves belonging to the same set of
pNSEES, NGAES , NARESq have identical markers. In Fig. 5.7 we see that the SEES scheme
outperforms GAES and ARES. For p1, 1, 1q, GAES performs better than ARES in the
depicted range, while in the other two configurations the two schemes perform almost
identical. Furthermore, the gap to SEES becomes very small.
83
Chapter 5 Channel Estimation
SEES SEES
GAES GAES
ARES ARES
−1
0
10 10
−1 −2
10 10
MSE
BER
−2 −3
10 10
−3 −4
10 10
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
|h| |h|
(a) (b)
Figure 5.8: 3 3 system, pNSEES , NGAES , NARES q p6, 7, 6q. (a) MSE and, (b) BER as a
function of the reference channel |h|.
Let the reference channels |h|, in this case all channel coefficients from the first transmit
antenna, be equal and take values in the range r0.001, . . . , 1s. For every fixed value
of |h|, the remaining channel coefficients hij are Gaussian distributed with σh2 1.
However, choosing small values of |h| induces a power penalty on the average channel
energy. In order to compare systems with equally strong channels on average, we
normalize the random channel coefficients with ε in (5.66). Thus, all H 1 have the same
average energy. Figures 5.8a and 5.8b depict the average Mean Square Error (MSE)
and BER, respectively, as a function of |h|. The MSE corresponds to E rkX x Xk2 s,
F
where X and X x are matrices containing the exact and the estimated sufficient CSI
X , respectively. Averaging is performed over the channel coefficients hij and over
noise. The performance of SEES and GAES are also plotted as reference. The SNR is
fixed at 25 dB. In this high SNR regime bit errors occur predominantly due to poor
channel estimation. As expected, the MSE of ARES is large for small values of |h|
and decreases as |h| increases. At the same time, the MSE of SEES suffers slightly for
very small |h| and reaches a constant value for all other values of |h|. GAES exhibits a
peculiar behavior since the MSE attains a minimum value close to |h| 0.25. The MSE
behavior does not translate directly to BER performance. In fact, all schemes perform
very poorly at low |h| and improve as |h| increases. ARES lags slightly behind the other
schemes, since it is affected more by small |h|. Note that in the region |h| P r0.05, 0.3s
the BER is poor despite the fact that the estimation MSE has reached the minimum
value for SEES. In fact, in this case the poor BER is due to noise, since small values of
84
5.1 Channel Estimation for Amplitude Detection
−1 −1
10 10
−2 −2
10 10
BER
BER
SEES (1,1,1)
GAES (1,1,1)
SEES (1,1,1) ARES (1,1,1)
GAES (1,1,1) GAES (1,2,2)
ARES (1,1,1) ARES (1,2,2)
−3 SEES (2,3,3) −3
10 10 SEES (2,4,5)
GAES (2,3,3) GAES (2,4,5)
ARES (2,3,3) ARES (2,4,5)
SEES (4,6,6) SEES (3,6,7)
GAES (4,6,6) GAES (3,6,7)
−4
ARES (4,6,6) −4
ARES (3,6,7)
10 10
−5 0 5 10 15 20 25 30 −10 −5 0 5 10 15 20 25
Average SNR (dB) Average SNR (dB)
(a) (b)
Figure 5.9: (a) Average BER for a 4 4 MIMO system, (b) Average BER for a 5 5 MIMO
system.
|h| reduce the decision regions, and hence lower the effective SNR. Finally, the impact
of small |h| on the BER of ARES is rather small compared to SEES and GAES.
44 MIMO: The parameters for this configuration are found in Table 5.1 for NT 4.
Fig. 5.9a depicts the average BER. Now we clearly observe that the systems which use
GAES and ARES for channel estimation do not exploit the full receive diversity —
which is two in this case — as does SEES (cf. Sectionsec:ch4SpatialMuxDiversityAD).
The performance of GAES is marginally better than ARES only at low SNR, and
GAES exhibits poorer diversity order.
55 MIMO: The parameters for this configuration are found in Table 5.1 for NT 5.
Fig. 5.9b shows that GAES has the poorest performance. Furthermore, the perfor-
mance saturates at high SNR, meaning that estimation errors are inevitable even when
the noise is low. In this case the performance of GAES is limited by error propagation,
as we will se next. ARES on the other hand avoids this issue and performs clearly
better than GAES. Once again, SEES exhibits the best performance. Notice that the
curves for p1, 1, 1q and p1, 2, 2q overlap. This is due to the fact that in the p1, 2, 2q
case, GAES and ARES have double the pilot length and half the power compared to
p1, 1, 1q, which in essence leads to almost identical performance.
GAES Error Propagation: As the number of transmit antennas increases, more
entries of the sufficient CSI are computed by iterative calls to cGAES. Additionally,
quantities computed by cGAES are used in new estimation steps as input arguments to
cGAES. This leads to a progressive degradation, since estimation errors are propagated.
85
Chapter 5 Channel Estimation
−2
10
MSE
−3
10
NGAES=1
NGAES=2
−4 NGAES=4
10
NGAES=6
1 2 3 4 5
Number of Summands
Figure 5.10: Error propagation in GAES: the MSE rises as larger sums are estimated.
This effect is quantified in Fig. 5.10. We use a 5 5 system, and depict the MSE of
the estimates:
$ { { { ,
& ¸3 ¸4 ¸5 .
%
|y
h11 |, |h11
{ h12 |
i1 i1 i1 -
, h1i , h1i , h1i , (5.67)
averaged over many channel realizations and for several values of NGAES . The first two
h11 | and |h11
estimates, |y { h12 |, are estimated directly using the corresponding pilot
A B
symbols from S and S in (5.38) and (5.39). The SNR is 25 dB, such that the
estimation quality of |y
h11 | and |h11 { h12 | is very good. The larger sums of three, four
and five coefficients are estimated through iterative calls of the cGAES algorithm. As
can be observed in Fig. 5.10, the MSE is almost identical for the first two estimates,
and degrades monotonically as further summands are added to the estimated norm.
h11 | and |h11
This behavior is independent of the estimation quality of the |y { h12 |, as
can be seen for various NGAES .
Conclusions: The above comparisons were made for almost equal pilot
length/spectral efficiency and for the same amount of energy invested in the train-
ing phase. We draw the following conclusions from the simulation results:
• For all antenna configurations and for a given pilot length5 , we observe that SEES
is always the best scheme in terms of performance.
SEES
• If the maximum allowed pilot length is less than Ntot for NSEES 1, then
5
This corresponds to comparing systems that are in the same row of Table 5.1.
86
5.2 Channel Estimation for Phase Detection
As we saw in Section 5.1 for amplitude detection, we only require knowledge of the
sufficient CSI X (cf. (5.4)) in order to perform ML detection. In the case of phase
estimation, it is not required to formulate the channel estimation problem along the
same lines.
We start our examination again with the ML detector that uses perfect CSI. We
rewrite (5.1) and (4.22)
¹
NR
spML argmax f py |z q argmax f pyi |zi q , (5.68)
s PS s PS
i 1
f pyi |zi q eρi ρi ρi sin2 ∆φi
e ?
cos ∆φi erfc p ρi cos ∆φi q , (5.69)
2π 4π
where ρi |zi|2{σw2 eTi Hs2 {σw2 and ∆φi yi =peT
i Hsq mod 2π P pπ, πs.
The dependence of the conditional pdf on the channel H takes place in ρi and ∆φi ,
through the norm and the phase of the hypothesis zi eT i Hs, respectively. In Sec-
tion 4.2.2 we discussed that the phase detector can distinguish between different phases
87
Chapter 5 Channel Estimation
and amplitudes, by utilizing knowledge of both |zi | and =pzi q (see Fig. 4.3a and 4.3b).
In the SISO case we can neglect the channel amplitude and only use the phase infor-
mation =pzi q in order to detect any M-PSK modulation. However, in the MIMO case
(see 4.2.3) this simplification is no longer possible. Knowledge of both |zi | and =pzi q is
required in this case. If we consider all modulation symbols of Ex-BPSK, we require
the norm and the phase of all linear combinations of the channel coefficients.
Hence, the MIMO phase detector requires complete channel knowledge. In Sec-
tion 5.1.1 the notion of sufficient CSI was motivated by the inability of the amplitude
detector to compute the complex-valued channel coefficients. However, it was possible
to estimate all hypotheses required by the ML detector, even without knowing the
channel matrix. In the case of phase detection it is possible to estimate the complex-
valued channel coefficients. For this reason, we will not use the notion of sufficient CSI
for phase detection.
First we investigate the scalar estimation problem. Estimating the phase shift of a
scalar channel using a phase detector seems straightforward, but estimating the ampli-
tude of the channel using phase samples is more involved. Both problems have been
covered in [13].
Assume the symbol s 1 is transmitted N times over the unknown wireless channel
h CN p0, σh2 q. The receiver observes
where |h| P R and φh P pπ, πs is the amplitude and phase of the unknown channel,
respectively. In the following, the two quantities will be estimated separately.
The phase φh can be estimated with a simple arithmetic mean. However, in order to
avoid the discontinuity that takes place at π rad, a vector estimator can be employed
instead [13]
¸
N
φph = exppjyi q . (5.71)
i 1
Estimating the unknown amplitude |h| is not straightforward. The key idea is that
the size of |h| affects the variance of the the phase observations at the receiver. For
88
5.2 Channel Estimation for Phase Detection
1
10
0
10
−1
10
σ2∆φ −2
10
−3
10
−4
10
−5
10 −2 −1 0 1 2
10 10 10 10 10
ρ1/2
example, large |h| implies high estimation SNR, meaning that the variance of the phase
samples will be low. On the other side, when |h| is small, the estimation SNR is low
and the same noise as previously leads to very high variance of the received phase
samples. This relation is used for estimating the amplitude |h|. Let the phase error be
∆φ rpy φh q mod 2π s P pπ, π s. The distribution of ∆φ, f p∆φ|ρq, is given in (5.69),
where ρ |h|2{σw2 . Note that ρ can be interpreted as the instantaneous estimation
SNR. The variance of ∆φ can then be computed as
»π
2
σ∆φ E ∆φ 2
∆φ2 f p∆φ|ρqd∆φ g pρq, (5.72)
π
as a function of ρ. Fig. 5.11 depicts (5.72) numerically computed. As mentioned
?
previously, as ρ |h| increases the variance of the estimation error decreases. Let
g 1pq be the inverse function of (5.72). An estimator for |h| is then given by
a b
|x
h| σw ρp σw g 1 σ
p∆φ
2
, (5.73)
and makes use of the estimated phase error variance (cf. [13]). The latter can be
computed as
¸
N 2 2
σ 2
p∆φ N 1 1
min yi p
φh , 2π
y i
φph . (5.74)
i 1
89
Chapter 5 Channel Estimation
1
10
0
10
N=10, 20, 40, 80
−1 MSE(h)
10
−2
MSE
10
−3
10
−4
10
−5 MSE(<h)
10
−10 0 10 20 30 40
SNR (dB)
p
h |x
h| ej φ .
p h
(5.75)
The estimator is based on the sample variance in (5.74), which implies that the number
of independent samples N, i.e. the length of the pilot, must be sufficiently large.
Furthermore, the estimator is not given in closed form, since the function g 1pq can
not be computed. However, this is not crucial for the complexity of the estimator,
since a look-up table for the inverse function can be used. Finally, note that the curve
?
for ρ p102 . . . 2 101 q in Fig. 5.11 is relatively flat, meaning that small errors in
the variance estimation will lead to significant errors in the estimation of ρ.
Fig. 5.12 depicts the MSE of the estimation of h CN p0, 1q. We depict the MSEs
E r|h p
h|2s and E r|φh φph q2 s for several values of N. The estimation of the phase
φh improves with SNR and with N. However, the MSE of h saturates as the SNR
increases. The reason for this behavior is that at high SNR, the size of N must be
very large in order to estimate the amplitude |h| correctly. As a consequence, the
estimation of h for any fixed N suffers from saturation. This is a limiting factor for
all estimation schemes that we present in the next Section. Summarizing the scalar
estimation problem, the estimation of the phase φh in (5.71) is simple and accurate
even for small N. On the other side, estimation of |h| in (5.73) and (5.74) is more
complicated, requires much larger N and the estimation performance saturates.
90
5.2 Channel Estimation for Phase Detection
In this section we develop three estimation schemes for MIMO phase detectors. The
fact that we can estimate the amplitude and the phase of the complex-valued channel
coefficients (cf. Section 5.2.1) suggests that we can directly estimate the channel ma-
trix. This direct scheme, called Direct Estimation Scheme (DES), is presented first.
However, as we will see, DES exhibits some shortcomings that necessitate the study of
two further schemes, SEES and ARES for phase detection. The latter are analogues
of the schemes developed for MIMO amplitude detectors in Section 5.1.
The DES estimates all entries of the channel matrix H using the scalar estimation
scheme from Section 5.2.1. The pilot of DES consists of NDES repetitions of the identity
matrix:
DES
Ntot NT NDES. (5.77)
Contrary to SEES for amplitude detection, DES does not suffer from the exponential
scaling of the pilot length with NT . However, as explained in Section 5.2.1, NDES must
be large in order to estimate the phase error variance in (5.74) reliably. DES suffers
from another effect however. As a consequence of the poor amplitude estimation, the
°
hypotheses zi eT i Hs j hij sj (cf. (5.69)) are erroneous even at high SNR. In
fact, both the amplitude |zi | and the phase =pzi q are erroneous, despite the fact that
estimation of phases is actually efficient with a phase detector (cf. (5.71)). Hence, both
the mean (=pzi q) and the variance of the pdf (5.69) are estimated poorly (cf. Fig. 2.2b).
As we will see in Section 5.2.3, this leads to poor estimates and consequently poor
performance at high SNR.
91
Chapter 5 Channel Estimation
As we saw, DES fails to estimate the phase of the hypotheses zi eTi Hs reliably, as
a consequence of the poor amplitude estimation. However, the phase of all required zi
could be estimated accurately if the pilot is chosen appropriately. This is the motivation
for applying the SEES for channel estimation using a phase detector. Recall that the
SEES (cf. Section 5.1.3.1) transmitted all symbols in S in the pilot, thus allowing the
receiver to estimate the sufficient CSI. In the case of phase detection, we similarly
train the receiver by transmitting all possible symbols. The receiver then estimates the
amplitude and phase of all hypotheses zi , without having to compute them indirectly
x
using a channel estimate H.
A special property exists when we consider Ex-BPSK. Since for every s0 P S,
s0 P S, it suffices to transmit only s0 in the pilot and estimate the corresponding
zi , i. The hypothesis for the case when s0 is transmitted, is readily obtained by
zi, i. This property reduces the length of the pilot by a factor two. Thus, consider
any partition of S in S 1 and S 2 , such that S S 1 Y S 2 and S 1 S 2 . The pilot of
SEES consists of all symbols s P S 1 . Each symbol is repeated NSEES times. The total
pilot length is
SEES
Ntot |S 1|NSEES 2N T 1 N
SEES . (5.78)
Similar to SEES for amplitude detection, the pilot length scales exponentially with NT .
However, due to the symmetry property of the symbol alphabet the total pilot length
remains short for small number of transmit antennas. SEES benefits from the fact that
the phase of the hypotheses zi is estimated directly and accurately. Good estimation of
the amplitudes |zi | is less important, and hence we expect to achieve good performance
for NSEES NDES .
A drawback of DES was that NDES must be very large in order to obtain good estimates
for the amplitudes of the channel coefficients. On the other side, the pilot length of
SEES grows exponentially with NT . These shortcoming motivate the development of a
scheme whose pilot length exhibits good scaling properties, and which does not require
large pilot length in order to estimate the amplitude of all channel coefficients.
These requirements are fulfilled by the last scheme we develop, ARES for phase
92
5.2 Channel Estimation for Phase Detection
ARES Pilot: The pilot sequence in ARES is composed of all symbols sARES P
S A Y S B Y S C Y S D defined as:
SA δer , (5.79)
SB tδe1 , . . . , δer1, δer 1, δeN u,
T
(5.80)
SC ts | s δper ek q for any k ru, (5.81)
SD ts | s δper jek q for any k ru, (5.82)
a
where δ pMARES 3pNT 1qNARES q{pMARES 5pNT 1qNARES q is chosen such
that every pilot symbol has average energy Es . Using MARES repetitions of S A , the
receivers estimates the channel coefficients hir . The remaining pilot symbols, S B , S C
and S D are repeated NARES times. Using S B , S C and S D we only perform phase
estimates, and thus MARES ¡ NARES or MARES " NARES . S C and S D have a similar
structure to ARES for amplitude detection. The total pilot length of ARES is
ARES
Ntot MARES 3pNT 1qNARES (5.83)
φij =z
hij , i, j r (5.84)
1
φir =h{
ir hij , i, j r (5.85)
2
φir =h{
ir jhij , i, j r. (5.86)
93
Chapter 5 Channel Estimation
ℑtu
h1 h2
h2
h1 jh2
h1
jh2
ℜtu
For the sake of simplicity, assume two random complex channel coefficients h1 and
h1 |y
h2 . We have an estimate p h1 |ejφ1 for h1 (obtained through S A ), and an estimate
φ2 for the phase of h2 (obtained through S B ). We furthermore have the estimates
φ3 =ph {h q and φ =ph{jh q (obtained through S C and S D , respectively).
1 2 4 1 2
p
hℑ
tan φ1 p
1
, (5.87)
hℜ1
hℑ
tan φ2 hℜ
2
, (5.88)
2
p
hℑ hℑ
tan φ3 p
1 2
, (5.89)
hℜ1 hℜ
2
tan φ1 tan φ3 p ℜ
p
hℜ
2 tan φ3 tan φ2 1
h , (5.90)
tan φ2 ptan φ1 tan φ3 q p ℜ
p
hℑ
2 tan φ3 tan φ2
h1 . (5.91)
The above solution is based on the tangents of all angles and on the real part of p h1 .
The tangents where preferred due to the direct relation to the real and imaginary parts
94
5.2 Channel Estimation for Phase Detection
of the vectors, without involving the hypothenuse. Note that φ4 is not deployed yet.
If we assume that all estimates are perfect, the above equations compute h2 perfectly.
Using (5.90) and (5.91), the estimate of h2 can be written as
2 2
p
h2 p ℜ
h2 jp
hℑ
jφ2
2e p
hℜ 2
p
hℑ2
ejφ2 . (5.92)
Similar to (5.63), we retain the accurately estimated phase φ2 in the final estimate.
However, there are certain cases where the computation of h2 is not possible using
only (5.90) and (5.91).
• h1 , h2 and h1 h2 are collinear. As a consequence, tan φ1 tan φ2 tan φ3 . In
this case, we compute h2 using φ4 instead of φ3 . Note that since h2 is rotated by
π
2
rad, tan φ4 tan φ1 and tan φ4 tan φ2. Using φ4 and the analogous equation
to (5.89), we compute:
tan φ4 tan φ1 p ℜ
p
hℜ
2 1 tan φ2 tan φ4 1
h , (5.93)
tan φ2 ptan φ4 tan φ1 q p ℜ
p
hℑ
2 1 tan φ2 tan φ4
h1 . (5.94)
95
Chapter 5 Channel Estimation
Table 5.2: Simulation parameters for DES, SEES, ARES and NT t2, 3, 4u.
NT pNSEES, NDES, NARES, MARESq Ntot
SEES GAES
Ntot ARES
Ntot
(5, 5, 1, 2) 10 10 10
(10, 10, 2, 4) 20 20 20
2
(15, 15, 2, 9) 30 30 30
(20, 20, 3, 11) 40 40 40
(5, 4, 2, 4) 15 16 16
(8, 6, 3, 6) 24 24 24
3
(10, 8, 3, 13) 30 32 31
(20, 15, 5, 30) 60 60 60
(4, 2, 1, 7) 15 16 16
(8, 4, 2, 14) 32 32 32
4
(12, 6, 3, 21) 48 48 48
(30, 15, 8, 48) 120 120 120
Hence, the case when h1 and h2 are collinear can be resolved using the alternate
solution given by (5.93) and (5.94).
• Tangents equal to infinity. When one of the angles equals π2 , the tangent equals
8. In that case, (5.90), (5.91), (5.93) and (5.94) can not be used. Furthermore,
if we consider the noisy case, small errors in the estimation of phases that are
close to π2 lead to very large errors in the tangents. These problems can be
removed by performing a pre-processing step. First, we rotate h1 , φ2 , φ3 and
φ4 by the same angle φp , such that no phase is close to π2 . Subsequently we
compute p h12 using (5.90), (5.91), or (5.93) and (5.94). Finally we counter-rotate
h12 back with φp and obtain p
p h2 .
In this section we examine the performance of DES, SEES and ARES by means of
system simulations. Table 5.2 summarizes the simulation parameters for NT p2, 3, 4q.
96
5.2 Channel Estimation for Phase Detection
DES
DES
−1 ARES
10 −1
10
Perfect CSI
−2 DES (5,5,1,2) −2
10 10
BER
BER
SEES (5,5,1,2) DES (5,4,2,4)
DES (10,10,2,4)
ARES (5,4,2,4) ARES
SEES (5,4,2,4)
ARES (10,10,2,4) DES (8,6,3,6)
SEES (10,10,2,4) ARES (8,6,3,6)
−3
10 DES (15,15,2,9) 10
−3 SEES (8,6,3,6)
ARES (15,15,2,9) DES (10,8,3,13)
SEES (15,15,2,9) ARES (10,8,3,13)
SEES (10,8,3,13)
DES (20,20,3,11) DES (20,15,5,30) SEES
ARES (20,20,3,11) ARES (20,15,5,30)
−4
SEES (20,20,3,11) −4
SEES (20,15,5,30)
10 10
−10 0 10 20 30 40 −10 0 10 20 30 40
Average SNR (dB) Average SNR (dB)
(a) (b)
Figure 5.14: (a) Average BER for a 2 2 MIMO system, (b) Average BER for a 3 3
MIMO system.
The parameters of all estimation schemes are chosen such that the total pilot length is
equal — whenever possible.
22 MIMO: Fig. 5.14a depicts the BER of 2 2 MIMO system with perfect CSI,
DES, ARES and SEES channel estimation. The parameters for channel estimation
are given in Table 5.2. We observe that the performance of DES is very good for low
and moderate SNR, where channel estimation is not critical, but starts to deteriorate
after a threshold SNR0 . This threshold is a function of NDES , and is higher for larger
NDES . Beyond this threshold, the estimation performance does not improve any more,
as can be seen in Fig. 5.12. However, the ML detector expects that the conditional pdf
(5.69) is increasingly concentrated around the mean. As a consequence, the detection
deteriorates as the SNR increases further. This behavior can be prohibited if the noise
variance σw2 in (5.69) is kept constant beyond the threshold SNR0 . The performance
of ARES is in general worse than DES for the same total pilot length, but ARES
only saturates at much higher SNR (with the exception of the p5, 5, 1, 2q setting, where
performance is very poor). Finally, SEES exhibits the best overall performance, and
remains close the the perfect CSI curve throughout the depicted SNR range.
33 MIMO: Fig. 5.14b depicts the BER of a 3 3 MIMO system. The obser-
vations in this case are similar as previously. DES reaches the saturation threshold
quickly, and performance starts deteriorating afterwards. In this setting, ARES also
saturates within the depicted SNR range. However, the performance of ARES does
not deteriorate afterwards. In fact, ARES estimates the relative phase between the
97
Chapter 5 Channel Estimation
DES
−1
SEES
10
−2
10 Perfect CSI
BER
channel coefficients and the reference coefficient. When computing the hypotheses zi ,
the estimated phase =pzi q does not suffer from the erroneous amplitude estimation of
the reference channel coefficient. This is the main difference between the performance
deterioration of DES and the saturation in ARES. Furthermore, ARES achieves lower
BER compared to DES in saturation. Finally, notice that the performance of SEES
also saturates, at least for the p5, 4, 2, 4q case. In total, SEES exhibits once again the
best overall performance.
44 MIMO: Finally we depict the BER of a 4 4 MIMO system in Fig. 5.15. The
performance is very similar to Fig. 5.14b. As the size of the MIMO system increases, it
becomes clear that all estimation schemes suffer from saturation. This unavoidable fact
is a consequence of the scalar estimation performance in Fig. 5.12. The main difference
between the various estimation schemes is the parameter N and the SNR where the
saturation begins. This implies that even the curves that do not suffer from saturation
will eventually saturate at some higher SNR.
Conclusions: The preceding simulation results revealed that channel estimation for
phase detectors suffers in general from the poor estimation of amplitudes using phase
samples. All three proposed schemes suffer from saturation of the BER at some SNR.
However, in terms of performance the SEES was a clear favorite. The performance was
actually very close to the case of perfect CSI. However, SEES exhibits an exponential
scaling of the total pilot length with the number of transmit antennas, a fact that limits
its relevance for large MIMO systems. In this case, the ARES or DES schemes could
be deployed instead, sacrificing however the performance significantly.
98
5.3 Conclusive Remarks
In this chapter we developed channel estimation schemes for MIMO amplitude and
phase detectors. In the case of MIMO amplitude detection, we showed that complete
channel knowledge is neither possible to obtain nor required by the ML detection.
We came up with the notion of sufficient CSI, and developed three schemes that can
estimate it. SEES, GAES and ARES differ significantly in the way they estimate the
channel, and yield a pilot whose length scales different with the number of transmit
antennas. In Section 5.1.4.2 we compared the three schemes and proposed guidelines
for choosing the appropriate scheme according to the scenario. In the next section
we will compare the three schemes in a real system. In the case of MIMO phase
detection, we saw that complete channel knowledge is actually required. We developed
three schemes, DES, SEES and ARES that compute either an estimate of the channel
matrix (DES and ARES) or the required sufficient CSI (SEES). Similar to MIMO
amplitude detection, the three schemes yield pilot lengths that scale different with the
number of transmit antennas. However, the estimation performance of all schemes was
eventually limited by the inability to perform accurate estimation of amplitudes using
phase samples.
99
Chapter 5 Channel Estimation
100
Part II
... to Practice.
101
Chapter 6
In this chapter we present the design of the sMILE test bed. The test bed is intended
for testing and verifying concepts that we developed in Part I of the thesis. This
chapter is structured as follows. First, we present an overview of the test bed and
its various components. The goals of the undertaking are clearly defined and a high
level description of the experiment is given. Next, we take a closer look at the various
building blocks: the design of the transmit signal, the custom-built sMILE nodes and
the software receiver. Finally, we conduct a series of measurements and characterize
the performance of the test bed for various scenarios. The measurements give further
insight into practical issues of MIMO amplitude detection. Appendix E presents further
details regarding the test bed, as well as various measurement plots, that have been
placed there in order not to disturb the text flow in this chapter. The contents of this
chapter will be published in [23].
6.1 Introduction
103
Chapter 6 sMILE: MIMO Amplitude Detection Test Bed
Hardware
Oscilloscope
RACooN sMILE
H
RACooN sMILE
Matlab
Figure 6.1: sMILE test bed hardware block diagram (here for a 2 2 MIMO system). Dashed
lines correspond to software ‘connections’.
development of hardware that is more power-efficient than existing designs, since this
is not possible in the framework of an experimental test bed. Such a task would require
System-on-Chip (SoC) development, which is outside of our expertise and the scope of
this thesis. We know however that amplitude detectors are in principle more power-
and cost-efficient compared to I/Q receivers, at the expense of performance.
In the following, we will deal with two types of effects:
• Generic: effects that origin from fundamental characteristics of our system, like
the use of amplitude detection and OOK modulation.
• Implementation-imposed: effects that depend on specific design decisions we
have taken, like using specific equipment or carrier frequency.
For serving the purpose of the test bed, it is important to identify and distinguish
between these two effects. In fact, the generic effects are more important, since they
give insight into the fundamental mechanisms that drive nonlinear MIMO systems.
The focus of our test bed is the nonlinear receiver. Design of the trans-
mitter is not crucial, so we resort to our lab equipment and choose to em-
ploy Radio Access with Cooperating Nodes (RACooN) nodes as transmitters. The
RACooNs are part of a 10-node cluster intended for implementing cooperative com-
munication protocols. RACooNs allow to transmit arbitrary baseband waveforms,
generated in Matlab, at carrier frequencies between 5.1 and 5.9 GHz. RACooN nodes
can transmit in a coordinated manner, acting this way as a virtual MIMO array. This
104
6.1 Introduction
Oscilloscope
RACooN sMILE
Initialize H
Figure 6.2: sMILE test bed software block diagram. The hardware part from Fig. 6.1 is
depicted in the middle.
is depicted in the hardware block diagram in Fig. 6.1. In this case, since each RACooN
employs its own clock reference, the configuration corresponds to a multiuser single-
antenna uplink transmission to a multiple-antenna receiver. However, it is also possible
to operate the RACooNs using the same clock reference. In this case, the configuration
corresponds to a single user multiple-antenna uplink. We will discuss the difference be-
tween the two configurations in Section 6.4.3. A short description of the RACooN lab
can be found in Appendix E.1.
The heart of the sMILE test bed is the amplitude detection MIMO receiver.
We design identical sMILE nodes, which individually perform amplitude detection.
The custom sMILE nodes implement only the RF front-end of the receiver, up to
the Analog to Digital Converter (A/D Converter), which is done by the oscilloscope
(LeCroy SDA6000). The recorded signals at the oscilloscope are then processed in
Matlab off-line. Although the receiver is implemented in a distributed way, we con-
sider this virtual MIMO system as a one-user, collocated MIMO array.
For our application, we use a symbol rate of fs 1 MHz, which is equal to the
per-antenna bit-rate, since we use binary OOK. The RACooN nodes have a signal
bandwidth of around 34 MHz. However, the choice of the RACooN nodes as trans-
mitters imposes important constraints to our design. In fact, in the 5 GHz frequency
band components are not aplenty, and circuit design and assembly is challenging. We
choose to build our test bed at a carrier frequency of 5.6 GHz, due to availability of
specific components for this frequency. The wireless channel at 5.6 GHz with 2 MHz
bandwidth is modeled as flat fading. Moreover, our experiment will be conducted in a
static lab environment such that slow fading can be also assumed.
Fig. 6.2 shows a block diagram of the software controlling the test bed operation.
The test bed is controlled exclusively through a Matlab simulator. The hardware part
of the test bed from Fig. 6.1 is shown in the middle of Fig. 6.2. A short description of
the high-level software blocks follows:
105
Chapter 6 sMILE: MIMO Amplitude Detection Test Bed
‘Initialize’ block: First we set various parameters for the simulator, like the size
NT NR of the MIMO system, the length of the synchronization sequence, the type of
the channel estimation scheme, pilot length, etc. A list of various parameters is given
in Table 6.1. Furthermore, we create NT independent information bitstreams stacked
in sdata P t0, 1uNT Ndata Npacket , which will be transmitted by the RACooNs.
‘Create Tx Signal’ block: Since we use OOK on every transmit antenna, the bit-
stream sdata readily corresponds to the mapped symbols. These are modulated with a
Nyquist pulse and we obtain the baseband transmit signal at the RACooN sampling
rate fRCN 80 MHz. Details regarding the design of the transmit signal are presented
in Section 6.2.
sMILE Hardware: Subsequently, the NT transmit signals are sent to the RACooN
nodes for the physical transmission. The nodes are first time-slot synchronized via
USB cable (cf. Appendix E.1), and then transmit their signals concurrently. The
digital oscilloscope captures the received signal on NR sMILE nodes. Details on the
hardware components of the testbed can be found in Section 6.3.1.
‘Read Oscilloscope’ block: The software simulator retrieves the noisy sampled and
quantized received signals from the oscilloscope over a GPIB bus.
‘Processing/Detector’ block: This block constitutes the main operation of the soft-
ware modem. The oscilloscope output is synchronized, the channel is estimated and
finally the transmitted bits are detected by the ML detector. The ‘Processing/Detector’
block is presented in detail in Section 6.4.
We proceed with the design of the transmit signal. Every transmission contains many
packets which constitute a burst. Fig. 6.3 depicts the structure of a packet. First, a
synchronization sequence of length Tsync is transmitted, followed by the pilot and the
data payload of duration Tpilot and Tdata , respectively. We choose a burst length of
Tburst 25.5 ms, almost the maximum length RACooN can handle. This corresponds
to 25, 500 symbols at 1 MHz symbol (bit) rate. The symbol/bit duration is Ts 1 µs.
106
6.2 Transmit Signal Design
Tburst
The packet length NSpP is a variable system parameter. If we choose NSpP 250, the
packet duration is Tpacket 250 µs and one burst consists of 102 packets. We assume
the channel remains constant throughout a packet duration. Table 6.1 summarizes the
various system parameters.
We use Ex-OOK, i.e., independent OOK streams on every transmit antenna. We use
a Nyquist pulse (Mth band filter) of total length 6Ts and roll-off factor β 1. Thus,
the baseband pulse bandwidth is 1 MHz. The pulse filter is designed with the Kaiser
window technique in order to achieve high side-lobe suppression. Fig. 6.4a depicts a
pulse at the sampling frequency of the RACooN baseband fRCN 80 MHz and the
corresponding frequency response.
107
Chapter 6 sMILE: MIMO Amplitude Detection Test Bed
40
1
20
0.8
0.6
Magnitude (dB)
−20
Amplitude
0.4
−40
0.2 −60
−80
0
−100
0 80 160 240 320 400 480 0 1 2
Samples Frequency [MHz]
(a) (b)
Figure 6.4: (a) Impulse response of Nyquist pulse at RACooN sampling rate, (b) and the
magnitude of the corresponding frequency response.
r1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1s. (6.2)
Both synchronization sequences are modulated to OOK symbols by mapping the ‘-1’
symbols into zeros.
Pilot Sequence As discussed in Section 5.1.4.2, we do not benefit from designing pilot
sequences with good autocorrelation properties. Instead, we resort to transmitting
repetitions of symbol s 1. Here, Npilot refers to the number of repetitions of a
training symbol, and is identical to the parameter N from Section 5.1.2. The exact
pilot structure varies according to the number of transmit antennas NT and the exact
channel estimation scheme used, i.e., for SEES, GAES and ARES the total pilot length
pilot
Ntot is given in (5.37), (5.40) and (5.51), respectively.
Fig. E.2 depicts the beginning of an exemplary transmit signal for NT 2. The
ramp-up phase of the Nyquist pulse is not shown, and the first synchronization bit is
depicted at t 0. The Barker code of length Nsync 11, given in (6.1), is followed by
the pilot sequence. In this case, Npilot 5. From t 11 µs to t 15 µs, Npilot ‘ones’
are transmitted from the first antenna, while the second antenna is silent (‘zeros’).
108
pq 6.3 sMILE Nodes
VCO
The sMILE nodes have been designed and assembled at our institute in cooperation
with Christoph Sulser. The most important component of each node is the amplitude
detector. In essence, this operation can be performed without the need of downcon-
version. The absence of the mixer and the LO would lead to tremendous power and
cost savings in such a design. However, we are constrained to implement our receiver
at 5.6 GHz. For practical reasons, we will first down-convert the received signal at an
Intermediate Frequency (IF), where a larger selection of more efficient components is
available.
Fig. 6.5 depicts an abstract block model of the sMILE nodes. The receiver has a
very straightforward structure. First, the received signal of fc 5.6 GHz at the an-
tenna is bandpass filtered at RF and then amplified by a Low Noise Amplifier (LNA).
The RF signal is subsequently mixed with a LO carrier at fLO , obtained by a
Voltage Controlled Oscillator (VCO). The resulting IF signal is first amplified and then
sharply filtered by a Surface Acoustic Wave (SAW) bandpass filter at fIF . Finally, the
filtered IF signal is fed to a Received Signal Strength Indicator (RSSI) detector, which
109
Chapter 6 sMILE: MIMO Amplitude Detection Test Bed
outputs the desired baseband signal Vout . The circuit could as well be implemented
only with an amplifier, a bandpass and an RSSI detector, if the carrier frequency was
lower.
All blocks in Fig. 6.5 are implemented with Surface Mounted Devices (SMD) com-
ponents. Table 6.2 lists the type and main characteristics of the various components.
The choice to operate the sMILE test bed at the carrier frequency of 5.6 GHz was
based on the availability of all components in Table 6.2 in the respective frequencies.
We use a free-running tunable LO for practical reasons. The oscillator frequency is
manually set at 5.2947 GHz. At this frequency, the 5.6 GHz signal is down-converted
exactly at fIF 305.3 MHz. Temperature variations and the lack of a PLL imply that
small frequency drifts will occur. To ensure that the desired signal will pass through
the IF passband filter, we chose the IF bandwidth WIF 12.5 MHz, wide enough to
account for LO drifts. By design, the signal at the input of the RSSI detector will
experience a net amplification of around 30 dB with respect to the signal in front of
the LNA. Finally, the only difference to designing a MIMO sMILE node is that a single
LO would be enough to feed the mixer on every receiver branch. Apart from that, no
other differences would occur to our distributed design.
The heart of the sMILE node is the RSSI detector. This component implements
amplitude/envelope detection, and corresponds to function g pq from our theoretical
system model in Fig. 2.1. The name is due to the most common use of such devices,
namely measuring the signal strength of a received signal. In its simplest form, an am-
110
6.3 sMILE Nodes
LNA
RSSI IF
Detector Bandpass RF
Bandpass
IF Amplifier
RSSI out RF in
(a) (b)
Figure 6.6: (a) RSSI Output vs. Input Level at 25 C for frequencies of 200 MHz, 300 MHz,
and 440 MHz [48], (b) Picture of sMILE node #0.
plitude detector can be implemented with a diode, a capacitor and a resistor (cf. [47]).
In practical systems, several Integrated Circuits (ICs) exist that implement amplitude
detection in a reliable and robust way. RSSIs are found either in the form of linear
amplifiers, or as so-called log-detectors. The latter achieve very high dynamic ranges
by means of a logarithmic amplifier. In fact, RF power measurement is a common ap-
plication of logarithmic amplifiers. For our test bed we choose the log-detector AD8310
with a nominal dynamic range of 95 dB (cf. [48]). The use of OOK necessitates high
dynamic range, since we need to distinguish between a high level, dependent on the
channel strength, and a low level which corresponds to no-signal/noise-only. Modeling
and understanding these devices is very important for the purpose of our test bed.
where Vin is the input voltage, Vout the output voltage and VY , VX are parameters. The
input-output characteristic of AD8310 is shown in Fig. 6.6a, as given in the manufac-
turer datasheet [48]. Notice that in the range r70 dBm, 15 dBms the response to a
300 MHz p fIF q input is linear. Hence, the device has a usable dynamic range of
85 dB for our application. For weaker input power levels, or for no input at all, the
111
Chapter 6 sMILE: MIMO Amplitude Detection Test Bed
output saturates around 400 mV. This corresponds to the noise floor. In this case,
the thermal noise at the device input is actually amplified and leads to the 400 mV
output. One technique to obtain a high dynamic range is progressive compression,
also used in AD8310. A cascade of six nonlinear amplifier cells is used to generate the
logarithmic function as a series of contiguous segments (piecewise linear technique).
A very weak signal is amplified by all six cells, while a strong signal is amplified by
the first cell(s) and drives the remaining cell(s) in saturation. Demodulation1 occurs
by nine detector cells. Six of them are associated with the amplifier stages, and three
are passive detectors that receive a progressively attenuated fraction of the full input.
The RF input signal, ranging from 0 up to 440 MHz, is demodulated and the output
signal is the envelope of the input, represented in a decibel scale. The output signal
has a so-called video bandwidth Wvideo 25 MHz. Since different signals traverse
through different number of amplifier cells, is is not possible to define a single noise
figure for log-amplifiers. Intuitively, the noise figure is high for weak input signals, that
are amplified by all amplifier cells, and low for strong input signals, that are amplified
by fewer amplifier cells. We will further study AD8310 in Section 6.3.2.
A photo of the prototype sMILE node #0 can be seen in Fig. 6.6b. The
Printed Circuit Board (PCB) is in-house designed and manufacturing is outsourced.
Soldiering was done in-house using a reflow oven. Subminiature Version A (SMA) con-
nectors carry the RF input and RSSI output signal. The microstrips are clearly visible
and the various components are labeled. The circuit is mounted in a metallic box
that offers both mechanical stability and isolation from electromagnetic interference
(though not high-level isolation). A total of 5 sMILE nodes have been built for the
purposes of our test bed. Each node consumes approximately 90 mA.
In this section we compute the Input/Output (I/O) characteristic of each sMILE node.
We will first compute the relation between the RSSI output voltage and the RF input
power, which is common practice for RSSI detectors. Then, we will relate the RF
power—given in dBm—to the input voltage, thus removing the logarithmic operation
of the RSSI detector. We are interested in the input voltage (in linear scale) since this
is the output of the theoretical system model from Part I of the thesis.
1
Here demodulation implies removal of the carrier from the signal.
112
6.3 sMILE Nodes
PinRSSI ptq
Vout ptq
VinRF ptq RSSI
PinRF
w ptq
LO
Since we do not have access to the input of the AD8310, we measure the I/O char-
acteristic of the whole sMILE node. Fig. 6.7 depicts a block model of a sMILE node,
where w ptq is AWGN added at the receiver front end. The output signal Vout ptq is sam-
pled by the oscilloscope. The input signal VinRF ptq A cosp2πfc tq is a cosine oscillation
at fc 5.6 GHz, generated with a signal generator and fed to the node input via cable.
Note that the reference impedance for all calculations is 50 Ω, the typical impedance
for wireless applications. The demodulated RSSI output is a Direct Current (dc) volt-
age whose value depends on A. Fig. 6.8 depicts the average output voltage in mV for
various input power levels, measured in dBm. Averaging is performed over the noise.
The sMILE nodes have almost identical I/O characteristic, with the exception of node
#1. This is probably due to the hardware assembly, which led to larger losses compared
to the other nodes. The input power level reaches only up to 23.5 dBm in order to
assure the components before the RSSI detector are not driven beyond the maximum
allowed power. Output voltages below 400 mV could not be measured reliably due
to very noisy RSSI output2 . Hence, the practical dynamic range3 is around 60 dB.
Notice that the voltage range in Fig. 6.8 is half the range in Fig.6.6a. This is due to
the circuitry at the output of the RSSI detector which acts as a 1{2 voltage divider.
The input power level (abscissa) discrepancy between Fig. 6.6a and Fig. 6.8 is due to
the RF front end, where the signal is mainly amplified (small attenuation occurs in
the bandpass filters and the mixer). Taking this into account, the power PinRSSI at the
RSSI input is on average 27.23 dB larger than PinRF, whereas sMILE node # 1 has the
lowest amplification. The missing 2.73 dB (on average) to the theoretical 30 dB net
amplification are lost due to mismatches and component tolerances.
As we can see in Fig. 6.8, the I/O characteristic is linear when the abscissa is in a dB
scale. In fact, the approximation of the logarithmic curve is done with great precision
2
Fig. E.3 in Appendix E.3 depicts the oscilloscope reading with and without input signal. The noise
when no signal is present has high variance.
3
The actual dynamic range is further reduced to 35 dB, as explained in Section 6.4.2.
113
Chapter 6 sMILE: MIMO Amplitude Detection Test Bed
1100
#0
1000 #1
#2
#3
900 #4
800
[mV]
700
out
V
600
500
400
300
−83.5−80 −70 −60 RF
−50 −40 −30 −23.5
Pin [dBm]
Figure 6.8: I/O characteristic of all sMILE nodes, measured with an unmodulated carrier
at the input.
by the AD8310. We approximate the I/O characteristic with a linear function, valid
in the range of measurement r83.5 dBm, 23.5 dBms
Vout f PinRF Vslope PinRF P0 rmVs. (6.4)
Using least squares fitting, we compute Vslope and P0 from the measured points of
Fig. 6.8. The parameter Vslope , calculated in mV{dB, describes the slope of the I/O
characteristic. P0 , measured in dBm, is the so-called log-intercept—the point where
the extrapolated linear response would intersect the abscissa. These parameters are
given in Table 6.3 for all sMILE nodes. For input power levels below 83.5 dBm the
output voltage eventually saturates close to a mean of around 200 mV, but the noise
of the measurement is so high that the reading is not reliable. In fact, signals below
400 mV are hard to distinguish visually from the noise level. As we will see later
however, received signals with a peak value around 400 rmVs can still be detected with
a BER of around 102 .
In Fig. 6.6a the abscissa is given in dBV, relative to 1 V Root Mean Squared (rms)
voltage of a sine wave, as well as in dBm. In reality logarithmic amplifiers react to
voltage only (and not to power), and hence, the I/O characteristic should always be
given in dBV. However, RF jargon usually deals with dBm, so the input power is
depicted as well. From the measured input power PinRF, we can compute the input
114
6.3 sMILE Nodes
10 log10 50 rpΩVsin 10
rVsq2 20 log |V rmsrVs|
rms
PinRF 3 rWs 10 in 13.01 rdBm{50Ωs. (6.5)
For the unmodulated carrier VinRF ptq A cosp2πfc tq we have Vinrms A2 {2. Thus, (6.5)
can be solved for the voltage |A|:
?
|A| 2 10αV rmVs β rVs
out
(6.6)
P0 13.01
α , β
1
, (6.7)
20V slope 20
where α and β are parameters, computed for every sMILE node separately using the val-
ues in Table 6.3. Note that in (6.6) we can only compute the absolute value/amplitude
of A. From (6.6) it is evident that in order to extract the input voltage, we have to ex-
ponentiate the RSSI output. This step removes the effect of the logarithmic amplifier.
The computed value corresponds to the peak voltage of a cosine wave at the input.
Hence, the input signal can be reconstructed according to:
?
VinRF ptq 2 10αVout rmVs β
cosp2πfctq rVs. (6.8)
?
Furthermore, notice that since the RSSI output is always positive, the peak voltage is
lower bounded by |A| ¥ 2 10β rVs.
where VinBB ptq is the baseband signal given in Fig. E.2. This transmit signal is char-
acterized by a Crest Factor (CF) of CFsignal 1.6749, which corresponds to the ratio
between the peak and the rms value. When the waveform in (6.9) is transmitted, the
output of the RSSI detector is a time-varying function that is proportional to the en-
115
Chapter 6 sMILE: MIMO Amplitude Detection Test Bed
1100
#0
#1
1000 #2
#3
900 #4
800
[mV]
700
out
V
600
500
400
300
−88 −80 −70 −60 −50 −40 −28
RF
P [dBm]
in
Figure 6.9: I/O characteristic of all sMILE nodes, measured with the signal from Fig. E.2 at
signal 4.48 dB with respect to Fig. 6.8.
the input. The curves are shifted by CFdB
velope of the received signal in logarithmic scale. The computed amplitude in (6.6)
now describes the norm of the baseband signal VinBB ptq, hence
BB ?
V
in pq
t 2 10αVout ptqrmVs β
. (6.10)
Notice that the output voltage in (6.10) is a time function. The reconstructed received
RF waveform at the sMILE node input is finally given by
?
VrinRF ptq 2 10αVout ptqrmVs β
cosp2πfctq, (6.11)
where the tilde is used to denote that we can not completely reconstruct the signal,
but only its envelope. If we now compute the power of (6.11), we will end up with
a different value than the one read from the I/O characteristic. This is because the
characterization was made using a different signal, namely an unmodulated carrier.
The correct value of the received RF power4 PsinRF is hence
Here, PinRF is the power obtained from the I/O characteristic using the peak of the
4
There exist RSSI detectors that take the CF of the signal into account and compute the true power
(e.g. ADL5505), but AD8310 is not such a device.
116
6.3 sMILE Nodes
replacemen
Tx Rx PinRSSI ptq
bk h1 ptq Vout ptq
gptq RSSI
pq VinRF ptq
w 1 ptq
LO LO
VinBB ptq
bk Vout ptq
g ptq hptq RSSI’
w ptq
Figure 6.10: Equivalent baseband system model.
time-varying envelope: maxt Vout ptq. The correction that follows considers the crest
factor of the reference (dc) signal and the actual signal used, where CFdB dc 0 and
CFsignal 20 log10 pCFsignal q 4.48 rdBs. Finally, the power of the signal in (6.11) is
dB
equal to PsinRF in (6.12). Hence, when the received signal at the sMILE node is the
signal from Fig. E.2, the I/O characteristic in Fig. 6.8 must be shifted by 4.48 dB to
the left, as shown in Fig. 6.9.
As a next step we develop a simplified baseband model that removes some complex-
ity from the RF front-end. Let us consider the system model in Fig. 6.10. The bit
sequence bk P t0, 1u modulates the transmit pulse g ptq, and the signal is subsequently
up-converted to the carrier frequency and transmitted. The received RF signal at the
receiver is VinRF ptq. In the lower part of Fig. 6.10 we draw the equivalent baseband
model, where hptq is the equivalent baseband channel and w ptq is the thermal noise
observed at the input of the RSSI detector. The new RSSI detector, denoted RSSI1 ,
models only the computation of the power in dB, as we will see in the next section.
That is, it does not include the demodulation functionality, i.e., removal of the carrier.
The baseband signal at the input of RSSI1 is VinBB ptq (see (6.9)). Using the equiva-
lent baseband representation, we have removed the carrier from (6.9), the mixers at
transmitter and receiver, as well as the two amplifiers and bandpass filters.
117
Chapter 6 sMILE: MIMO Amplitude Detection Test Bed
pq
VinBB ptq
1
t nTos
bk
Vpout rns
g ptq hptq |
| 1
α
log10 ?210β
Figure 6.11: Modeling of RSSI1 in the equivalent baseband representation. Sampling time
1 20 ns.
Tos fos
Using the equivalent baseband model from Fig. 6.10 we proceed with the noise char-
acterization of the sMILE nodes. Fig. 6.11 depicts the equivalent baseband model,
where the RSSI1 is substituted with the idealized logarithmic operation (excluding the
demodulation). At the input of RSSI1 , the AWGN w ptq which describes the baseband
thermal noise that has traversed through the receiver front-end is added to the sig-
nal. The RSSI detector generates the additive noise ν ptq, which is depicted in Fig. 6.11
inside the RSSI1 block. The RSSI detector is then sampled at the oscilloscope at a sam-
pling frequency of fos 50 MS{s. During quantization at the A/D Converter of the
oscilloscope, the quantization noise q rns q pnTos q is added on the discrete-time signal,
where n is the discrete time index. Finally, we obtain the discrete-time quantized RSSI
output voltage:
We denote noisy signals with a hat in the following. With the Nyqust rate of fos ,
we sample the 25 MHz wide RSSI output without aliasing. The signal of interest is
concentrated at a much smaller bandwidth, and hence fos is high enough, meaning that
no Intersymbol Interference (ISI) occurs. Using the noisy oscilloscope measurement, we
compute the noisy baseband voltage VpinBB rns using (6.10) and (6.13) as follows
?
in
V nr s 2 10αVp rns β
p BB
out
(6.14)
?
2 10αpV rns νrns qrnsq
out β
(6.15)
VinBBrns 10αpνrns qrnsq . (6.16)
118
6.3 sMILE Nodes
¸
VinBB rns bk xrn kTs {Tos s w rns, (6.17)
k
where xrns g rns hrns. Hence, the inferred noisy baseband voltage can be written as
¸
αpν rns q rnsq
rs
p BB
Vin n
r
bk x n kTs
s r s
w n 10
loooooomoooooon (6.18)
loooooooooooooomoooooooooooooon
k Ψ
Φ
In (6.18) we can see the effect of all three noise sources on the finally computed baseband
signal envelope. We observe two components, Φ and Ψ , which describe two different
mechanisms.
If we neglect the RSSI and quantization noise, i.e., Ψ 1, the resulting signal is
identical to the theoretical system model (cf. Section 2.2 for SISO). In this case, the
desired signal is obscured by the AWGN samples w rns, and we observe the absolute
?
value of the superposition. The distribution of the signal is expected to be Rician,
except for a bias5 equal to |Vin,min
BB
| 2 10β . Hence, Φ describes the fundamental
behavior of our model, which is a result of amplitude detection.
However, the RSSI noise ν rns and the quantization noise q rns can not be neglected.
The combined noise Ψ is multiplied with the component Φ, and the effect is similar to
that of a fading channel on the transmitted symbol. First, we take a look at the two
noise components and their impact on the RSSI output Vout , as described in (6.13).
The quantization noise q rns is generated at the oscilloscope’s A/D Converter, which
performs a uniform 8-bit quantization of the RSSI output. q rns is thus uniformly dis-
tributed, with variance σq2 ∆2 {2, where ∆ is the smallest quantization level (cf. [49]).
The RSSI generated noise ν rns depends on the strength of the input signal. Small
signals propagate through many amplifier cells inside the RSSI detector, and the re-
sulting ν rns has high variance σν2 , while strong signals are amplified less and lead to
small σν2 . When observing Vout rns, the signal is noisy when it is weak, and less noisy
when it is strong. The quantization noise is uniform throughout the range of the signal
strength, that is, the quantization levels ∆ are constant. Fig. E.3 in Appendix E.3
depicts Vpout rns without any input signal, and when an unmodulated carrier is trans-
mitted. By inspection we can see that the noise variance is very large when no signal
is received, and significantly less for a certain received signal strength.
The above effects change, however, when we observe the computed input voltage
5
This shift stems from the fact that the RSSI detector is lower bounded by the noise floor, when the
input signal is absent.
119
Chapter 6 sMILE: MIMO Amplitude Detection Test Bed
p BB
rs
Vin n . We distinguish two cases:
• Strong input
signal
Φ: When the input signal is strong we can neglect ν rns. In
this case, Vin rns Φ 10αqrns . The product with a large Φ leads to amplification
p BB
|V | Φ 10αq (6.19)
|V 1| Φ 10αq1 Φ 10αpq ∆q (6.20)
∆|V | |V | |V 1 | Φ p10α∆ 1q. (6.21)
influence on Vpout rns. The different effects are a result of using a log-detector and
sampling the logarithmic output with a uniform quantizer. Both using a log-detector
and the oscilloscope which quantizes uniformly are design decisions that were imposed
by practical constraints. Hence, these effects are not fundamental to the nonlinear
MIMO system that we study, but rather implementation imposed. In Section 6.4.2 we
6
In this and similar figures following: large square, star and circle markers stand for synchronization,
pilot and data bits, respectively. The norm of the transmit signal is drawn with a dashed line.
Round empty markers depict the sampled points at symbol rate. The instant t 0 corresponds
to the first synchronization bit of the first received packet. The transmit and received signals are
plotted in different scales, depicted left and right of the plot, respectively. The plot is drawn after
synchronization is done.
120
replacemen 6.4 Software Receiver
Channel
Synchronization Nsync Estimation X
perform a statistical analysis of the noise behavior and quantify the above observations.
So far we have described all components of the sMILE test bed, except for the ‘Pro-
cessing/Detector’ block of Fig. 6.2. This block handles the digital processing of the
signal sampled by the oscilloscope, and includes the basic functions of synchronization,
channel estimation and ML detection. Fig. 6.12 depicts a block diagram of the various
processing cells. The vectors in Fig 6.12 correspond to the signal received by NR sMILE
nodes, and acquired by NR oscilloscope channels. In most parts of the software receiver
the different received streams will be processed independently. However, MIMO ML
Detection is performed jointly using all streams simultaneously. The various processing
cells are described below:
rns (see
‘Exp’ block: First, the noisy oscilloscope output Vpout (6.13)) is used to com-
pute the norm of the noisy baseband input voltage Vinrns. This processing is per-
p
formed on each receiver branch separately, using (6.14). The values of the parameters α
and β correspond to the respective sMILE node where the signal was actually received
(see 6.7 and Table 6.3). Fig. E.4 depicts an exemplary oscilloscope output (SISO case)
and the corresponding noisy baseband input voltage. Until t 4995 µs only noise is
captured at the oscilloscope. Notice the shift from the logarithmic to the
linear
do-
main: the noise-only part is reduced to very small values close to zero in Vpinrns. The
notches in Vpout rns are around the noise floor at approximately 0.2 V, while the same
notches in Vpin rns almost reach zero.
121
Chapter 6 sMILE: MIMO Amplitude Detection Test Bed
‘LP’ block: As we saw in Fig. E.5 the positions where ‘ones’ are transmitted are
very noisy. In order to smoothen
the
signal and reduce the effect of the exponentiated
quantization noise, we filter Vpin rns using a low pass filter. Filtering is performed with
an FIR digital filter with 3 dB bandwidth at 4 MHz. The filtered version |Vpinf rns| is
subsequently shifted in order to be synchronous with the unfiltered version, i.e., the
group-delay of the low-pass filter
is removed.
Fig E.6 depicts the filtered version of the
received signal from Fig. E.5. Vin rns is now much smoother and less noisy. However,
pf
the filtering operation is done heuristically and leads to degradation of the points where
‘zeros’ are expected. In fact, the ‘zeros’ are elevated in Fig. E.6 compared to Fig. E.5.
In the following we will keep both the filtered and unfiltered version of the signal, as
shown in Fig. 6.12. We will use the unfiltered version in order to detect ‘zeros’ and the
filtered version in order to detect ‘ones’. This is rather uncommon, but nevertheless
leads to improved performance of our system.
The output of the synchronization block is a matrix Nsync that contains the estimated
sampling points for all bits across all receivers. Every packet is synchronized separately.
The synchronization algorithm is described in Section 6.4.1 in detail.
‘Sampler 1 MHz’ block: This block down-samples the signals from the oscilloscope
sampling frequency fos 50 MS{s to the symbol rate fs 1 MHz. We use the
estimated sampling points stored in Nsync by the synchronization block.
Both
the
filtered and unfiltered signals are then down-sampled, resulting in Vinrms and Vin rms,
p pf
respectively. The index m distinguishes the sequences at fs from the sequences at fos .
less noisy when observing ‘ones’. The output of the channel estimation process is the
sufficient CSI X tX1 , . . . , XNpacket u for all packets in a burst (see 5.1.1). We assume
that within a packet the channel remains constant, but may change from packet to
packet. Channel estimation is described in Section 6.4.3 in detail. However, in order
to estimate the channel we must know the distribution of ‘ones’. This is handled in
Section 6.4.2 together with the distribution of ‘zeros’.
122
6.4 Software Receiver
‘ML
Detection’
block: Finally, we use the estimated sufficient CSI X , |Vpinrms| and
Vin rms and perform MIMO ML Detection. In order to compute the required metrics
pf
we need the distribution of ‘ones’ and ‘zeros’. These will be computed in Section 6.4.2.
The ‘ML Detection’ block is described in Section 6.4.4 in detail.
6.4.1 Synchronization
Synchronization is performed separately for each packet and independent for each
receiver branch/oscilloscope channel. Since all three sequences Vpout rns, |Vpinf rns| and
|Vpinrns| are synchronous, we can estimate the synchronization point using any of the
three sequences. The use of |Vpinf rns| leads to inferior performance, since the ‘zeros’
of
the signal are altered after low-pass filtering (see ‘LP’ block). Using Vout rns or Vinrns
p p
is preferable, and between the two, Vpout rns leads to slightly better performance. The
synchronization process is handled as follows.
Subsequent packets: The rest of the packets are synchronized in the same fashion.
However, since the beginning of the first packet has been estimated, we know the
position of the following packets more accurately. Hence, the correlation search for the
rest of the packets is constrained to only a few samples around the assumed packet
beginning.
Finally, all estimated sampling points for all received branches are stored in Nsync
and passed to the ‘Sampler 1 MHz’ block.
123
Chapter 6 sMILE: MIMO Amplitude Detection Test Bed
transmitted at the ith position in the SISO case, or if the all-zero symbol was sent
in the MIMO case, the receiver(s) samples noise-only. Hence, the sample from the
unfiltered sequence |Vpinris| should
be used.
If any non-zero symbol is transmitted, the
sample of the filtered sequence Vin ris should be used, since it experiences less noise
pf
statistics across all sMILE nodes in order to examine whether different nodes have
different behavior.
Statistics of ‘ones’: Figures E.7a, E.7b and E.7c depict histograms of y1 rms for
three different signal levels. In all three cases, the data are very well approximated by
a Gaussian distribution. In Fig. E.7a we also see a fitted Rician distribution, which
is actually indistinguishable from the Gaussian distribution (cf. 2.2). For the sake of
simplicity, we use the Gaussian distribution henceforth. The mean µ1 of the Gaussian
distribution is a function of the signal strength, as expected. The variance σ12 of the
124
6.4 Software Receiver
Gaussian distribution also depends on the signal strength, and grows when the signal
strength is higher. This is in line with the observations made in Section 6.3.4, i.e., that
the noise is higher for high signal values. The mean of the Gaussian distribution is an
estimate of the channel strength |h|. In fact, a simplified system model for y1 rms is
y1 rms |h w 1 | |h| w, where w N p0, σ12q is real Gaussian noise. Hence, the
mean of y1 rms is the ML estimate |x h|. In Fig. E.8a we plot the estimated standard
h| µ1 . As
deviation of the Gaussian distribution for all sMILE nodes as a function of |x
expected, the standard deviation σ1 increases as |h| grows. All sMILE nodes7 have a
x
similar behavior. As can be seen however, the curves are not monotonous, but instead
exhibit stationary points. To the best of our understanding, these occur whenever the
RSSI detector switches a further amplifier cell on or off. Unfortunately, it is impossible
to account for this behavior. In Fig. E.8b we perform a linear fitting for all sMILE
nodes and obtain a relation between the standard deviation and the channel strength:
! ) ! )
µ1 , σ1 || ||
x
h ||
h , 0.004483 x
x h 0.0026 . (6.22)
The Gaussian pdf and (6.22) will be used by the ‘ML Detection’ block in Section 6.4.4.
Statistics of ‘zeros’: Figures E.9a, E.9b and E.9c depict histograms of y0 rms for
three different signal levels. In all three cases, the data are best fitted by a lognormal
distribution. In Fig. E.9a we can also see how a Rayleigh fitting is not suitable for
y0 rms . Hence, the ‘real’ distribution of ‘zeros’ is not Rayleigh, as expected by theory.
On the other side, the noise samples as seen in Fig. E.3 and E.4 are seemingly Gaus-
sian distributed. So it is no surprise that after exponentiation in the ‘Exp’ block the
sampled ‘zeros’ are lognormal distributed. The parameters µ0 and σ0 of the lognormal
distribution are estimated for different signal levels and sMILE nodes in Fig.E.10a.
The abscissa is the mean of the ‘ones’ for each transmission, that corresponds to the
estimated channel strength |xh|. We observe that the variance σ0 is constant for all
sMILE nodes and all |h|. The mean µ0 on the other side is weakly dependent on |h|
x
in a linear fashion. In Fig. E.10b we compute the average of all σ0 values, as well
as the fitted linear relation between µ0 and |h|. Summarizing, the parameters of the
lognormal distribution that describes y0 rms are
! ) ! )
µ0 || x
h , σ0 | | 5.555, 0.4341
0.478 x
h . (6.23)
7
Here, as well as in Fig. E.10a, we have omitted sMILE node #1, which exhibits different behavior
with respect to the other nodes. sMILE #1 will not be used henceforth.
125
Chapter 6 sMILE: MIMO Amplitude Detection Test Bed
The lognormal pdf and its parameter will be used by the ‘ML Detection’ block in
Section 6.4.4.
Note that the statistics computed above are valid for fixed settings at the oscillo-
scope. That is, whenever the amplification/zoom level at the oscilloscope changes, the
quantization noise affects the signal in a different range. The result is a horizontal shift
of the curves in Figures E.8a, E.8b, E.10a and E.10b. We chose to conduct all mea-
surements and operate the test bed within the window r0, 0.8 Vs on the oscilloscope.
This choice was dictated by operational properties of the oscilloscope. Unfortunately,
constricting the sMILE output voltage in the above range leads to a reduction of the
dynamic range to approximately 35 dB (cf. Fig. 6.9).
As we saw in Section 6.4.2, the ‘ones’ of the received signal are Gaussian distributed.
This makes the process of channel estimation very easy. All channel estimation schemes
discussed in Section 5.1 break down the estimation problem in several scalar estima-
tions. The scalar estimator used was either the MoM estimator, or a combination
of the MoM and the Bayesian estimator (cf. Section 5.1.2). However, since we now
observe a Gaussian distribution instead of a Rician, the scalar estimator, discussed in
Section 5.1.2, is as simple as the sample mean. The rest of the estimation protocols
and the corresponding observations from Section 5.1 remain the same—only the scalar
estimation step is now different. The channel estimation is performed for every packet
in the received burst.
Assume we want to estimate the sufficient CSI Xp of the pth packet. Consider the
case when sm P S is the training symbol, which was sent at the mth position of the
pilot (cf. Section 6.2). The corresponding entry in the sufficient CSI can be estimated
as follows:
¸
Nsync mNpilot
xsm N 1
pf
rs
Vin k . (6.24)
pilot
k Nsync
p q
m 1 Npilot 1
Here we locate the part in the filtered baseband signal which corresponds to reception
of the pilot symbols associated with sm , which follows after Nsync synchronization bits.
We then compute the numerical average of these samples and estimate the mean of
the Gaussian pdf. This step is performed simultaneously across all receiver branches.
126
6.4 Software Receiver
User 1 User
RACooN RACooN
sMILE LO
sMILE
User 2
receiver receiver
RACooN RACooN
(a) (b)
Figure 6.13: Racoons in a (a) single-antenna multiuser configuration, (b) single-user mul-
tiple antenna configuration.
If we use SEES, then all entries of Xp are estimated with (6.24). If GAES or ARES
is employed, then (6.24) covers only a part of Xp . The remaining entries are then
calculated as described in Sections 5.1.3.2 and 5.1.3.4.
At the end of the pilot the symbol s3 r1, 1sT is repeated Npilot times. Using (6.24),
we compute a numerical average that corresponds to an estimate for |h11 { h12 | and
|h21{h22 |. We use two configurations for the RACooNs. First, the RACooN nodes
operate independently as shown in Fig. 6.13a. The LOs of the two nodes exhibit a
variable phase difference due to independent phase noise. This corresponds to two
single-antenna users that transmit in a coordinated fashion to a multiple antenna re-
ceiver. Then, we connect the two RACooN nodes such that they share the same LO, as
shown in Fig. 6.13b. This implies that the LO phase noise between the two transmitted
signal is identical. This is the case of a single user that employs two antennas.
Multiuser scenario: Fig. E.11 depicts the output of such a channel estimation for
a 2 2 system with Npilot 10. In this case transmission was conducted via cables
using RF splitters/combiners and the RACooNs were employed as shown in Fig. 6.13a.
127
Chapter 6 sMILE: MIMO Amplitude Detection Test Bed
The estimates |y
hij | of the direct channel coefficients are not affected by phase noise on
the LOs, since only one node transmits each time and the phase of the transmitted
signal is removed at the receiver. Furthermore the estimates are practically constant
throughout the whole burst, since transmission is conducted over a static cable channel.
The estimates of the sums |h11
{ h12 | and |h21
{ h22 | on the other side change between
packets8 , since they are affected by different phase noise between the LOs of the two
transmitting RACooNs. Assume the phase noise of the two LOs is φ1 ptq and φ2 ptq,
respectively. In reality, the receiver computes estimates of
|h11ejφ ptq
1
h12 ejφ2 ptq | |h11 h12 ej∆φptq |, (6.27)
|h21ejφ ptq1
h22 ejφ2 ptq | |h21 h22 ej∆φptq | (6.28)
where ∆φptq φ2 ptq φ1 ptq is a random variable describing the difference of the phase
noise processes on the two LOs. The phase noise ∆φptq can take values up to 20
among consecutive packets (cf. [50]). This effect is always present whenever multiple
independent users transmit, since each one uses a separate LO. As a consequence,
frequent estimation of the channel is required.
Figures E.12a and E.12b depict the estimated sufficient CSI during two wireless
transmissions. The position of transmitters and receivers was not changed between
the two transmissions and the RACooNs were again employed as shown in Fig. 6.13a.
However, the Line-of-sight (LoS) path between the first transmit antenna and the sec-
ond receive antenna was blocked by an object in the second transmission (Fig. E.12b).
Compared to Fig. E.11, the estimates |y hij | change more dynamically throughout the
burst, and in some cases exhibit oscillations. This is due to the variable wireless propa-
gation environment, affected among others by the fluorescent lights in the room. Once
again, the biggest channel changes between neighboring packets are in estimation of
|h11{h12 | and |h21{h22 |, which are affected by phase noise (cf. (6.27) and (6.28)).
h21 | is significantly weaker in Fig. E.12b, since this LoS path is
Finally, the coefficient |y
obstructed. The remaining coefficients |y hij | |y
h21 | are relatively unchanged between
the two transmissions.
Single-user scenario: Finally, we estimate the sufficient CSI when the RACooNs
use the same LO reference, as shown in Fig. 6.13b. The estimated sufficient CSI is
depicted in Figures E.13a and E.13b for a wireless and wired transmission, respectively.
8
In Fig. E.11 this effect is visible only in |h11
{h12 |.
128
6.4 Software Receiver
Contrary to Figures E.12a and E.12b, the estimates |h11{ h12 | and |h21
{ h22 | do not
suffer from different phase noise, since both transmitted signals exhibit the same phase
noise perturbation, that is removed by the RSSI detectors.
Example 4 4: Figures E.14a, E.14b, E.14c and E.14d depict the sufficient CSI of a
4 4 configuration. This time, RACooN #1 and #2, and RACooN #3 and #4 employ
the same LO, respectively. Hence, the four RACooNs form two virtual users employing
two antennas each. This mixed scenario combines all effects we discussed for the 2 2
case. The pilot for NT 4 and SEES is given below (for Npilot 1)
1 0 0 0 1 1 1 0 0 0 1 1 1 0 1
0 1 0 0 1 0 0 1 1 0 1 0 1 1 1
sSEES
pilot .
(6.29)
0 0 1 0 0 1 0 1 0 1 1 1 0 1 1
0 0 0 1 0 0 1 0 1 1 0 1 1 1 1
The sufficient CSI consists of p2NT 1qNR 60 entries. First, the 16 direct channel
hij |, i, j are depicted in Fig. E.14a. These correspond to the first four
estimates |y
pilot symbols in (6.29) and are fairly constant throughout a burst. Fig. E.14b depicts
estimates of sums of all combinations of two channel coefficients. These correspond
to columns 5 10 in (6.29). The four thick lines and the four thick dashed lines
correspond to columns 6 and 10 of the pilot, respectively. These are the cases where
only RACooN #1 and #2, or RACooN #3 and #4 are transmitting. Since these
pairs of transmitters use the same LO, the estimates do not suffer from different phase
noise. The remaining (thin) curves in Fig. E.14b correspond to transmitters that use
different LOs, and hence the estimates change more dynamically. Fig. E.14c depicts
the estimates of sums of three coefficients, that correspond to columns 11 15 in
(6.29). Now, at least one transmitter uses a separate LO, and hence all estimates
are prone to phase noise. Finally, Fig. E.14d depicts the four estimates when all four
RACooNs transmit, which corresponds to the last pilot symbol in (6.29). Again, all
four estimates are prone to phase noise. Notice however that the estimates do not
always exhibit dynamic variations as a consequence of different phase noise, as can be
seen in the cyan curve in Fig. E.14d.
129
Chapter 6 sMILE: MIMO Amplitude Detection Test Bed
6.4.4 ML Detection
ML Detection is the final step in the sMILE test bed. The estimated sufficient CSI
is used together with the down-sampled filtered and unfiltered sequences in order to
detect the data bits. The output of the ML detection block is the detected bit sequence
and the respective BER.
Assume we want to detect the ith transmitted symbol in the pth packet. The suffi-
cient CSI for the pth packet, Xp has been computed by the ‘Channel Estimation’ block.
The ML Detector picks the symbol spris that maximizes the likelihood function Λi psm q
for all possible transmitted symbols sm P S in the symbol alphabet:
In the upper branch of (6.31) we obtain the product of NR lognormal pdfs. The
argument of each pdf is the jth element (jth receive antenna) of the unfiltered signal
vector |Vpinris| at the ith position. In the lower branch of (6.31) we obtain the product
of Gaussian pdfs, and the argument is this time the jth element of the filtered signal
vector |Vpinf ris| at the ith position. The individual pdfs are computed as follows:
f0 |eT p
j Vin r s|
i sm 1
? epln |e TVp
j in ris|µ0 pxsj qq2 {2σ02 , (6.32)
|eTj Vpinris|σ0 2π
f1 |eTj Vpinf ris| sm 1
? ep|e TVpf ris|xj,sm q2 {2σ12 pxj,sm q . (6.33)
σ1 pxj,s q 2π
j in
The parameters tµ0 pxsj q, σ0 u of the lognormal distribution in (6.32) have been computed
in Section 6.4.2. There we used |h| as the argument of µ0 . In the general MIMO case
it is not clear what the argument of µ0 should be. We use x sj , which is the average
130
6.5 Operation and Measurements
of the sufficient CSI X at the jth receive antenna for all possible sm . An alternative
would be to use the average of µ0 in Fig. E.10b for all |h|, since the dependence on
|h| is week. When we make the assumption that a non-zero symbol was transmitted,
all receiver branches receive a non-zero signal level. Hence, the pdf is the Gaussian in
(6.33) with mean equal to the expected level xj,sm . This is defined as the entry in Xp
that corresponds to the case when sm is transmitted and to the jth receive antenna,
i.e. the jth element of (6.24). The respective standard deviation σ1 pxj,sm q is given in
(6.22).
Example 4 4: Fig. E.15 depicts the transmit signal of a 4 4 MIMO system. The
first 75 µs of the four transmit streams are shown. The Nsync 11 synchronization
pilot
bits of the Barker code are followed by Ntot 60 pilot bits. The SEES method is
used with Npilot 4. For t 11 . . . 14 µs only the first transmitter sends Npilot 4
‘ones’ and the remaining transmitters are silent. The rest of the pilot is described in
Section 5.1.3.1. The random data bits follow after t 71 µs. Fig. E.16 depicts the
corresponding received signal on all four sMILE nodes. The first 75 µs are shown.
From the overlapping of received and transmitted signals during the synchronization
sequence (t 0 . . . 10 µs) we conclude that synchronization of the first packet was
successful. During reception of the pilot sequence, marked with black stars, one can
observe the various levels that correspond to all possible transmit symbols. In this
131
Chapter 6 sMILE: MIMO Amplitude Detection Test Bed
−1
−1 10
10
BER
BER
−2 −2
10 10
(a) (b)
Figure 6.14: (a) BER of a SISO transmission over cable, as a function of the input RF power
Psin
RF (cf. (6.12)). Measured with sMILE #0, (b) Simulated BER of OOK in
case, the different amplitude levels—that correspond to entries of the sufficient CSI—
are well separable. This leads to good performance. In fact, the BER over one burst in
this example was zero. Fig. E.17 depicts t 79 . . . 100 µs of the same received signal.
This part of the signal corresponds to reception of random data bits. At t 92 µs the
transmitter sends the all zero symbol s 0. The received signal on all four receivers
goes down to zero. In this case, all four received samples are lognormal distributed,
and the likelihood of the received sample vector is given in (6.31) and (6.32). All other
transmitted symbols in Fig. E.17 contain at least a single non-zero entry. Hence, the
samples on all four receivers are different from zero, and the corresponding likelihood
function is computed using (6.31) and (6.33).
BER and Noise Figure: In order to characterize the hardware of the sMILE receivers
we perform a BER measurement of a SISO configuration using a cable for transmission.
By means of a variable attenuator we measure the BER for various input power levels.
The BER curve is depicted in Fig. 6.14a and was measured with sMILE #0. The
abscissa is the computed input RF power PsinRF. We used a 31 symbol long m-sequence,
and Npilot 20 pilot length in order to achieve reliable synchronization and channel
estimation. Thus, bit errors occur mainly due to the noise realization. Note that the
measured points in Fig. 6.14a do not represent an average over many transmissions.
Fig. 6.14b depicts the simulated BER of OOK in an AWGN channel, that has been
computed in Section 4.2.1. We now take a BER of 102 as reference. This BER
132
6.5 Operation and Measurements
occurs for PsinRF 89.3 dBm and corresponds to an SNR of 8.3 dB. Hence, the
noise floor at the RF input is Pnoise PsinRF SNR 98.2 dBm. This inferred noise
power includes the noise amplification by all hardware components. The theoretically
computed noise power at the RF input, considering the IF bandwidth of 12.5 MHz and
the aggregate noise figure NF 3.7 of all components except for the RSSI detector,
theory
equals Pnoise 99.3 dBm. Hence, the inferred realistic noise floor Pnoise is only
1.1 dB off the theoretical value. If we further assume the the measurements are error-
free, this 1.1 dB can be interpreted as the noise figure of the RSSI detector at the
given operating point. Note however that for reasons discussed in Section 6.3.1 it is
not possible to characterize the RSSI detector with a single noise figure.
Typical parameters: Table 6.4 summarizes typical values for the operation and per-
formance of the test bed. The number of users, shown in the third column, corresponds
to the way the RACooNs share a LO reference: if the number of users is less than the
transmit antennas NT , then two RACooNs use the same LO and form a virtual single
user with two transmit antennas (cf. Section 6.4.3 and Figures 6.13a, 6.13b). The last
column states the spectral efficiency of the system, computed according to
Eq. (6.34) considers the number of bits per symbol, NT (due to the use of Ex-OOK), the
efficiency loss due to synchronization and estimation symbols, and finally the symbol
rate fs and the passband bandwidth Wpb (cf. Table 6.1). During all measurements
the RACooN output power was set to "LOW", which corresponds to approximately
10.5 dBm (cf. Appendix E.2). However, for the two measurements marked with a star
the output power was set to "HIGH", which corresponds to approximately 7.5 dBm.
The various measurements in Table 6.4 are commented below:
1. Synchronization: The first two measurements are conducted via cable in order
to attain a static channel. We compare the impact on the BER of the 11-Barker
code and pm 31q-sequence, in low SNR conditions. The received signal power
at the sMILE node is 94 dBm. The long m-sequence leads as expected to better
BER. The erroneous bits are 3852 when the Barker code is used, and 3241 when
using the m-sequence.
133
Chapter 6 sMILE: MIMO Amplitude Detection Test Bed
Environment
Propagation
distance [m]
Synchroniz.
Estimation
[bits/s/Hz]
Efficiency
Sequence
NT NR
Channel
Spectral
Scheme
Npilot
Users
NSpP
BER
11 1 - Barker 250 SEES 20 0.172 0.438.
➀
11 1 - m-sequ. 250 SEES 20 0.159 0.398
21 6.1 102
cable
1 - Barker 250 SEES 10 0.756
➁
22 1 - Barker 250 SEES 10 0 0.756
11 1 20 Barker 250 SEES 10 103 0.458
11 1 37 Barker 250 SEES 10 0 0.458
2.5 102
➂
22 1 25 Barker 250 SEES 10 0.836
22 1 37 Barker 250 SEES 10 2.5 104 0.836
33 2 10 Barker 150 SEES 6 3.3 103 0.97
➃ 33 2 10 Barker 150 GAES 7 9.5 103 0.97
33 2 10 Barker 150 ARES 6 0 0.97
44 1.2 102
LOS
2 10 Barker 500 ARES 3 1.8360
6 105
➄
44 2 10 Barker 80 ARES 3 0.9750
44 2 10 Barker 150 SEES 2 2.4 103 1.4533
44 2 10 Barker 150 GAES 3 3.2 102 1.4533
44 2 10 Barker 150 ARES 3 5.5 103 1.4533
1.1 103
➅
44 2 10 Barker 150 SEES 4 1.0533
44 2 10 Barker 150 GAES 6 1.3 102 1.0533
44 2 10 Barker 150 ARES 6 5 103 1.0533
44 2 - Barker 200 SEES 4 5.5 103 1.29
1.2 102
➆
44
NLOS
4 - Barker 200 SEES 4 1.29
➇ 44 2 mixed - Barker 200 SEES 4 1.1 103 1.29
Fig. E.18 depicts the received signal in a 2 2 constellation. The received signal
on the first antenna is around 21 dB weaker than on the second antenna. Note
that the different levels during the channel estimation phase are well separable on
the first antenna. As a consequence, it is possible to distinguish between the four
different transmitted symbols using only the first received signal, as if we had a
2 1 Multiple Input Single Output (MISO) system. The resulting BER equals
6.1 102, and is a result of poor signal to noise ratio and detecting a virtual
pM 4qOOK. If the received signal on the second antenna is also employed
during ML detection, as in a regular 2 2 system, the BER is reduced to zero.
3. Range: The following four measurements demonstrate the range of the test bed.
The measurement environment is sketched in Fig. E.20a. At the maximum length
of the corridor, d 37 m, the RACooN’s transmit power is set to 7.5 dBm
("HIGH"). The resulting BER is still very low. Since it is possible to further
134
6.5 Operation and Measurements
increase the RACooN transmit power by 16 dB, the range exceeds the maximum
corridor length. For the maximum RACooN transmit power of 23.5 dBm (cf. Ap-
pendix E.2) and a target BER of 102 in the SISO case, the maximum path loss
is 112.5 dB. Using a path loss exponent of 3, the theoretical range is then 150 m.
4. 33 channel estimation: The three 3 3 measurements compare the performance
of SEES, GAES and ARES with values for Npilot taken from Table 5.1. For the
specific choice of Npilot , all three systems have the same spectral efficiency. The
BER for all three estimation schemes is low. However, since we deal with wireless
transmissions, it is not possible to recreate exactly the same conditions for all
three transmissions. Hence, the BERs should not be considered as absolute
figures of merit.
5. Packet length: The following two measurements demonstrate the effect of the
packet length on the BER. We compare two packet sizes NSpP 500 and
NSpP 80. All other parameters are identical. When very short packets are
used the channel is estimated very often. Hence, the estimated sufficient CSI is
‘always’ up-to-date and the BER is very low. However, this compromises the
spectral efficiency which in this example is reduced to only one half. When the
packet length is long, the sufficient CSI becomes outdated during the duration of
the packet, and as a consequence the ML detector fails to detect appropriately.
In fact, the BER is dramatically higher for NSpP 500, indicating that outdated
CSI is a major source of bit errors.
6. 44 channel estimation: Like in the 3 3 case (4th measurement), the following
measurements compare the three channel estimation schemes for values of Npilot
that result to identical spectral efficiencies. The same observations hold in the
4 4 case as well, although this time GAES has the tendency to perform poorer
than the other schemes. Recall that as NT grows, the performance of GAES
deteriorates (cf. Section 5.1.4.2).
7. Number of users: The following two measurements are conducted in
Non Line-of-sight (NLoS) conditions, which reduce the received power signifi-
cantly. The measurement environment is sketched in Fig. E.20b. We compare
two 4 4 systems with different number of users. The case of four/two users
corresponds to the extension of Fig. 6.13a/6.13b to four transmitters. The BER
in the two-user case is better than in the four-user case. This is due to the impact
of LO phase noise on the channel estimation, as explained in Section 6.4.3. The
effect is similar to the impact of the packet length (see 5th point). Again, the
135
Chapter 6 sMILE: MIMO Amplitude Detection Test Bed
errors occur due to outdated CSI. A way to improve the BER is reducing the
packet length.
We described the design and operation of the sMILE test bed. To the best of our
knowledge, the test bed is the first implementation of a communication system that
enables detection of spatially multiplexed streams using envelope detectors. The re-
ceiver is based on RSSI detectors, although such devices are not used for detection
purposes in practice. This emphasizes the low complexity and cost attached to such
type of receivers, especially when implemented in a SoC.
Throughout the design of the test bed, we were confronted with issues that originated
from the component availability and the equipment involved:
• The use of a logarithmic amplifier-type RSSI detector implied dealing with a de-
vice whose noise amplification is a function of the input power (cf. Section 6.3.1).
136
6.6 Conclusive Remarks
137
Chapter 6 sMILE: MIMO Amplitude Detection Test Bed
138
Chapter 7
139
Chapter 7 Summary and Conclusions
performing ML detection. Once again, certain differences were revealed between am-
plitude and phase detection. In the case of amplitude detection, we came up with
the notion of sufficient CSI and developed three channel estimation schemes. SEES,
GAES and ARES resulted from three fundamentally different views of the estimation
problem, and have significant differences in terms of performance and pilot length scal-
ing. Each one is suitable for different scenarios. Estimating the channel for a MIMO
phase detector proved to be more challenging, since in this case complete knowledge
of the complex-valued channel coefficients was required. We developed three further
channel estimation schemes for the phase detector: DES, SEES and ARES. Similar to
the case of amplitude detection, the three schemes have significant differences in terms
of performance and pilot length scaling. However, all three schemes suffer from the
inability of the phase detector to perform accurate estimation of amplitudes.
Concluding our work on nonlinear MIMO systems, we built the sMILE test bed.
This is the first MIMO amplitude detection test bed, and demonstrates that such a
system works in principle. We were able to implement the three channel estimation
schemes we developed in theory, and obtained valuable insights into the design and
operation of such systems.
This thesis offered a first exploration of nonlinear MIMO systems. Capitalizing on
our findings, we conclude that nonlinear MIMO systems are a promising new design
direction for low cost/low power systems. The deployment of multiple antennas in such
systems enables performance gains that have been thus far only possible with linear
MIMO systems and their respective linear I/Q receivers. As the need for ever cheaper
and low-power devices constantly increases, nonlinear MIMO should be consider as a
viable solution.
Since this was the first consideration of nonlinear MIMO systems, the topic could
not be investigated exhaustively. Several open questions are worth answering, shedding
more light into the world of nonlinear MIMO systems and their applications. During
the development of the sMILE testbed we identified several areas that require further
consideration. For example, a straightforward extension would be coded transmis-
sion, which requires identifying and adapting coding schemes to the needs of nonlinear
MIMO systems. While our hardware design was constrained by the available equip-
ment and components, it is worthwhile considering non-uniform quantizers and other
envelope-detection specialized components. This implies using the existing state of the
art designs for amplitude detection in a MIMO implementation. Another interesting
topic is exploring how nonlinear MIMO nodes would operate in a cooperative communi-
140
cations scenario, where they would act as relays. This relay functionality is especially
interesting if we consider the ubiquitous nature of such low cost/low power devices.
Although we considered both MIMO amplitude and phase detection throughout this
thesis, our attention was mainly in amplitude detection. However, phase detection is
fundamentally more powerful and should be further investigated. Finally, in order to
quantify the value of such systems, a SoC implementation of a sensor node employing
nonlinear MIMO would emphasize the power/performance gains compared to existing
designs. In combination with compact MIMO antennas, this would probably give a
major boost to the adoption of nonlinear MIMO systems in future sensor networks.
141
Chapter 7 Summary and Conclusions
142
A Monte Carlo Integration
Monte Carlo integration is an efficient method for computing multidimensional inte-
grals. Compared to traditional numerical integration techniques, Monte Carlo esti-
mates converge to the correct value at a rate that is not dependent on the dimensional-
ity. An extended version of the following brief introduction can be found in [51]. First,
we see how numerical techniques suffer from the so-called curse of dimensionality.
Suppose we wish to compute the integral
»
I f pxqdx, (A.1)
Ω
¸
N
Ip wi f pxi q, (A.2)
i 1
where the weights wi and the sample points xi are determined in advance. The weights
can be computed with various methods, like the Newton-Cotes rule or the Gauss-
Legendre rule, for one-dimensional integrals. In the one-dimensional case the above
methods are very efficient, provided the integrand has sufficiently many derivatives.
The convergence rate is typically of the form O pN r q, for some integer r ¤ 1. Now,
consider the following multidimensional expression:
¸
n ¸
n ¸
n
Ip wi1 wiD f pxi1 , . . . , xiD q. (A.3)
i1 1 i2 1 iD 1
This method has the same convergence rate as the one-dimensional case, e.g., O pnr q.
However, the number of sample points is much larger: N nD . Hence, the conver-
gence in terms of N is only O pN r{D q, and is a function of the dimensionality D. The
efficiency of computing (A.3) diminishes as the dimensionality grows, an effect called
143
A Monte Carlo Integration
the curse of dimensionality. An equivalent point of view is that the volume of a hy-
percube grows exponentially with additional dimensions. Thus, sampling a hypercube
with fixed sample length requires exponentially more samples as the dimensions grow.
Monte Carlo integration is a means to avoid the above mentioned problem. Suppose
we wish to approximate (A.1). Assume we possess N independent samples Xi of a pdf
ppxq. An approximation to (A.1) is given by
1 ¸ f pXi q
N
Ipmc N i1 ppXi q
. (A.4)
The choice of the pdf ppxq is important. It must have the same support as Ω, and we
must be able to create random samples distributed according to ppxq. If Xi are chosen
independent and uniformly distributed in Ω, then
¸
N
Ipmc N1 f pXi q, (A.5)
i 1
which is similar to (A.2), except that the sample points are now randomly distributed.
The estimate of the integral is unbiased.
1 ¸ f pXi q
N
E Ipmc E
N i1 ppXi q
(A.6)
»
f pxq
Ņ
N
1
ppxq
ppxqdx (A.7)
» i1
Ω
f pxqdx (A.8)
Ω
I. (A.9)
1 ¸
N
Ipmc N i1
Yi . (A.10)
144
For any Yi ,
»
f 2 pxq
var rYi s E Yi2 E rYis
2
dx I 2
ppxq
(A.11)
Ω
we assume that var rYi s is finite. Now, we compute the variance of Ipmc :
¸N ¸N ¸
N
var Ipmc var N1 Yi 1
N2
var Yi N12 var rYi s
1
N
var rYi s . (A.12)
i1 i1
i 1
145
A Monte Carlo Integration
146
B Diversity-Multiplexing Tradeoff
for SISO Amplitude/Phase
Detection
We consider s is Gaussian distributed (cf. [19]) with variance σs2 . For a given channel
realization, hs w is Gaussian distributed CN p0, σw2 |h|2σs2q, and the amplitude y|h
is Rayleigh distributed according to
f py |hq
y y22
e 2λ , λ 2
σw2 |h|2σs2 . (B.2)
λ2 2
The entropy of a Rayleigh distributed random variable is known, and given by ( [52])
hpy |hq γ λ
1 ln (B.3)
2 2
1
γ 1 σw2
ln
|h|2σs2 rnatss, (B.4)
2 2 4
147
B Diversity-Multiplexing Tradeoff for SISO Amplitude/Phase Detection
like
y y2 2σ|hs|2 y |hs| 2
f py |s, hq e 2
r I0 , σr2 σ2w . (B.5)
σr2 σr2
We can approximate1 the Rice distribution for very low values of σr2 with a Gaussian
distribution of variance 2σr2 σw2 . The entropy is then given by
hpy |s, hq rnatss.
1
ln 2πeσw2 (B.6)
2
From eqs (B.1), (B.3), (B.6), the instantaneous rate at high SNR (or low σw2 ) is ap-
proximately equal to
2
?
where κ 1 γ
2
ln 8πe for brevity, and SNR σσ 2
s
2
w
. Eq. (B.8) is similar to linear
detection. The difference is the constant k, and the square root inside the logarithm.
As we will see next, the square root will affect the multiplexing gain.
We now compute the probability that the outage rate is less than r ln SNR for high
SNR:
P |h|
2
2κ
1
e SNR 12r
1
SNR
P |h|2 1
e SNR12r
2κ
(B.9)
1
e2κ SNR1 2r
. (B.10)
The approximation in Eq. (B.9) holds due to the high SNR assumption. In the last step,
we used the fact that h is Rayleigh distributed and the approximation Pp|h|2 ǫq ǫ,
for small ǫ [3]. Consequently, the diversity-multiplexing tradeoff is given by [3]
ln Pout pr ln SNRq
dAD pr q lim 1 2r. (B.11)
SNRÑ8 ln SNR
1
The Rice distribution is approximated by a Gaussian when the Rice factor K |hs|2{p2σr2 q Ñ 8.
148
B.2 Phase Detection
We proceed in a similar way for deriving the diverisity-multiplexing tradeoff for phase
detection. However, we restrict the input distribution to the following: f psq ejφ ,
where the phase φ is uniformly distributed in pπ, πs. Consequently, the SNR is
defined as SNR σ12 . First we show that y |h
w
y argpz q argp|h|ej pφ p qq
arg h
w q argp|h|ejθ wq (B.12)
is also uniformly distributed in pπ, πs. Here, the auxiliary random variable θ is
uniformly distributed in pπ, π s, which follows from circular symmetry and the fact
that h |h|ej argphq is considered deterministic. It suffices to show that z is circularly
symmetric. For this, z must satisfy two conditions: E rz s 0 and E rz 2 s 0. z can be
rewritten as
wr |h| cospθq j pwi |h| sinpθqq,
z (B.13)
where wr , wi N p0, σw2 {2q. The first condition follows immediately from Eq. (B.13).
For the second condition, we have
0. (B.14)
Hence, z is circularly symmetric. Since the phase of z takes values in the whole range
pπ, πs, y argpzq is uniformly distributed in pπ, πs. The entropy of y|h is conse-
quently given by
hpy |hq lnp2π q rnatss. (B.15)
149
B Diversity-Multiplexing Tradeoff for SISO Amplitude/Phase Detection
where ρ |hs|2{σw2 |h|2{σw2 and ∆φ y argphsq P r0, 2π q. In the high SNR regime,
the distribution can be simplified using the following approximations
∆φ ! 1 ñ sin ∆φ ∆φ (B.17)
cos ∆φ 1 (B.18)
?
ρ Ñ 0 ñ erfcp ρ cosp∆φqq 2 (B.19)
eρ
2π
0 (B.20)
σ2
y |s, h N argphsq, w 2
2|h|
. (B.21)
at high SNR. Again, the structure of Rphq is similar to linear detection, while the
main difference is the existence of the square root. Finally, the instantaneous rate is
computed from Eqs. (B.1), (B.15) and (B.22) as
d d
4π |h|2
Rphq lnp2π q ln
σw2 2πe
2|h|2
ln eσw2
4π |h|2
ln
e
SNR. (B.23)
150
B.3 Conclusions
dpr q
Linear detection
1{2 1
r R{ ln SNR
Figure B.1: Diversity-multiplexing tradeoff
The approximation in Eq. (B.24) is justified like in Eq. (B.10). Finally, the diveristy-
multiplexing tradeoff is given by
ln Pout pr ln SNRq
dPD pr q lim 1 2r, (B.25)
SNRÑ8 ln SNR
B.3 Conclusions
Fig. B.1 depicts the diveristy-multiplexing tradeoff of amplitude, phase, linear detection
and BPSK. The endpoint dp0q 1 describes the classical diversity gain of amplitude
and phase detection, which is the same as for linear detection. The endpoint dprmax q
0 ñ rmax 12 provides analytical proof of our numerical result, that the multiplexing
gain of amplitude and phase detection is half that of linear detection. As a matter of
fact, the diversity-multiplexing tradeoff of amplitude and phase detection is the same
as that of BPSK modulation. In both cases, the system “looses” half the degrees of
freedom, albeit in different ways.
151
B Diversity-Multiplexing Tradeoff for SISO Amplitude/Phase Detection
152
C Energy of Extended OOK
NT
NT 1
1 ¸ NT α2 ¸ NT 1
Es 2NT k1 k
NT
kα2 2NT k0 k
NT α2 NT 1
2 NT
2 α2 N2T , (C.1)
a
where we used identity (3.1.6) from [34]. Finally, we obtain α 2Es {NT .
153
C Energy of Extended OOK
154
D Geometrically Assisted
Estimation Scheme (GAES)
F
0.8
0.6 h3
G
h3
0.4
E
h +h
1 3
h2+h3
0.2
h3 B
h2
h3
0 O
O h1+h2
δ h2
φ H D
−0.2 12,3
φ1,3 φ1,3
D h
1
φ A
A 2,3
φ1,2
−0.4
−0.8 −0.6 −0.4 −0.2 0 0.2
155
D Geometrically Assisted Estimation Scheme (GAES)
D
φ
2,3
0.8
θ
φ
1,3
h
2 φ12,3 h3
0.6
A G
H
φ
1,2
h1+h2
0.4 h φ1,3 B
3
h h
1 2
E
0.2 h3
δ
h1+h3
h +h
F
0 2 3
O
h3
−0.2
C
−0.4
−1 −0.5 0 0.5 1 1.5 2 2.5
0.5
A
h2
φ θ
1,3
h
φ
1,2 D
1
φ12,3
δ h +h
δ φ12,3
0 O θ
1 2
H φ2,3
h2 B
h +h
1 3
−0.5
h h3
3
h h3
3
−1
h2+h3
E
G
−1.5
C
F
−1.5 −1 −0.5 0 0.5 1 1.5
156
D.1 GAES Topologies
1.8
1.6
1.4
1.2 G
h
3 φ1,3
1 φ2,3 D
E
0.8 θ δ H
h2 φ H
12,3
0.6 h3 h2+h3 h
3
h1+h2
h1+h3
0.4 A
h1 δ h3 B
0.2 h2
0
3.5
G
h3
D
φ
3 12,3
φ
2,3
H δ θ h1+h2+h3
2.5 h
2 F
h3
B
2
h2 h1+h2 h2+h3
1.5
h E
3
1 A
δ
h1
h +h
1 3
0.5
C
h
3
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
157
D Geometrically Assisted Estimation Scheme (GAES)
Further measures are necessary in order to allow cGAES to operate with noisy esti-
mates. First, the computations in lines two and three in Algorithm 1 must be modified
such that the argument of arccospq is bound in r1, 1s, hence resulting in valid real
angles in r0, π s. The extreme cases of angles equal to zero or π correspond to degen-
erated triangles, where the edges are collinear. This degenerations are frequent at low
SNR, where the estimates from S A and S B are very noisy, but actually occur at all
SNR, since the channel coefficients are random variables. In this case, cGAES detects
the occurrence of such a case and computes |h1 h2 h3 | accordingly. E.g., when
φ1,2 φ2,3 φ1,3 π, ~h1 κ~h2 λ~h3 , with κ, λ ¡ 0 (in fact, ~h1 , ~h2 and ~h3 are ‘close’
to collinear) and consequently φ12,3 π. Similarly, φ1,2 φ1,3 π, φ2,3 0 means
~h1 κ~h2 λ~h3 , with κ, λ 0 and subsequently φ12,3 0. Other degenerate cases are
handled in a similar way.
Finally, special handling is necessary for two more cases when △OAD is degenerate.
In both cases, φ1,2 π and θ δ 0. First, if φ2,3 π and φ1,3 t0, π u, we
are in case A2.2. However, applying cGAES in this case yields φ12,3 π, which is a
course approximation that ignores φ1,3 . Instead, we do the approximations θ1 δ 1
pπ φ1,3q{2 and φ12,3 π θ1 . Finally, if tφ2,3 , φ1,3u t0, πu, we compute the angle of
interest as follows: φ12,3 minpφ1,3 , φ2,3q θ1 , and θ1 |φ1,3 φ2,3 |{2.
158
E sMILE Test Bed
The RACooN is a custom-built radio test bed at our institute. It comprises 10 identi-
cal single-antenna nodes that can transmit and receive at carrier frequencies between
5.1 GHz and 5.9 GHz in a half-duplex manner. Fig. E.19 depicts two clusters of
RACooN nodes stacked on top of each other. Each node employs a power supply unit
(PSU), a storage unit (STU, baseband processing and data storage) and an RF unit
(RFU, 16 bit AD/DA converters, IF/RF mixing). Individual Rubidium normals are
used to provide the clock for each unit. The baseband clock rate is 80 MHz leading
to a symbol duration of 12.5 ns. The nodes exhibit user bandwidth with equalized
phase response of 60 MHz. Timeslot synchronization takes place using USB cables
prior to operation. Each RACooN node is placed on a cart to provide mobility. We
use short dipole antennas that have been designed for the required frequency range
and are mounted on poles at a height of about 1.5 m. The RACooNs can be oper-
ated through Matlab. The transmit signal is provided to the RACooNs as a complex
baseband signal, whose amplitude ranges between 1 to 1. This signal is up-converted
to the desired carrier frequency and transmitted. The output power of the RACooNs
can be set to "LOW" or "HIGH", which corresponds to 7 dBm and 25 dBm,
respectively, when the baseband signal is a real sine wave with amplitude 1. Further
details on the RACooN lab can be found in [50].
The RACooN nodes have a side-effect that severely influences the sMILE test bed.
Before the signal reaches the mixer for up-conversion (transmitter mode), a dc voltage
is leaked on the baseband signal. This dc component can not be controlled, and after
up-conversion, it results in a carrier that is sent along with the intended RF signal. This
159
E sMILE Test Bed
interfering carrier has a power of approximately 25 dBm when the transmit power
is set to "LOW", and 7 dBm when se to "HIGH". If a RACooN is used for data
reception, the interfering carrier results in a dc component in baseband. This can be
easily removed in post-processing. However, when using a sMILE node as the receiver,
the interfering carrier is demodulated along with the intended signal. For example, if
the transmit power is set to "LOW" and we account for a minimum path loss of 40 dB,
the interfering carrier has a power of 65 dBm at the sMILE node front-end. This
is well above the noise floor, that lies at 98.2 dBm (cf. Section 6.5). Thus, when a
zero is transmitted, the output voltage of the sMILE node does not reach the noise
floor, but instead is raised to a level above the noise floor. This level varies according
to the path loss between transmitter and receiver. In essence, the carrier is interfering
with our signal and ‘destroys’ the transmitted zeros. Unfortunately, there is no way to
remove the carrier either at the transmitter nor at the receiver.
In order to cope with this shortcoming and remove the effect of the carrier leakage
we use a trick at the transmitter side. We make use of the 34 MHz available baseband
bandwidth and up-convert our signal to an intermediate frequency fIF 1 20 MHz.
Our new baseband signal occupies the range r19 MHz, 21 MHzs, and the dc leakage
is 20 MHz away from the new center-frequency. Next, the new baseband signal is
mixed to a new carrier frequency fc1 fc fIF 1 5, 58 GHz. The intended signal is
still centered at fc 5.6 GHz. Hence, the receiver will shift the signal such, that the
frequency fc will fall in the center of the IF bandpass filter. The intended signal will
pass through the filter, and the interfering carrier will lie 20 MHz away from the IF
center frequency. Fig. E.1 depicts the frequency response of the IF filter as given in the
datasheet. As we can see, the attenuation at 20 MHz away from the center frequency
is between 30 dB and 40 dB with respect to the passband signal.
Finally, we operate the RACooN nodes at "LOW" transmit power and furthermore
add a 10 dB attenuator in front of the transmit antenna. Hence, the leaking carrier
is attenuated by a further 30 dB to 40 dB, and reaches the receiver at 115 dBm
to 105 dBm (these computations are summarized in Table E.1). This is well below
the noise floor, and hence the carrier does not alter the received signal. However,
we have constrained the transmit power significantly. Taking the crest factor of our
signal into account, the mixing to IF, the effective use of a single sideband and the
10 dB attenuator, the RACooN transmit power is 10.5 dBm when set to "LOW",
and 7.5 dBm when set to "HIGH". If we chose to keep the interfering carrier, we would
have to train the receiver for the case when a zero symbol is transmitted. This way we
160
E.2 Dealing with RACooN’s DC leakage
Figure E.1: Frequency response of IF bandpass filter RF3602D between 255 MHz and
355 MHz.
would estimate the interference level and account for it in ML detection. We did not
follow this path in order to have a test bed that is as close as possible to the theoretical
system model from Part I of the thesis. As explained in Section 6.4.2, the oscilloscope
operates within a certain range and the maximum received signal power is 55 dBm.
Within this range, the interfering carrier is always below the noise floor.
Note that the range of the test bed does not necessarily suffer by this artificial
transmit power reduction. When the distance and hence the path loss are high enough,
the RACooNs can gradually send with higher transmit power and even without mixing
to IF . In this case, the path loss attenuates the signal enough such that the carrier is
below the noise floor.
161
E sMILE Test Bed
0.8
Amplitude
0.6
0.4
Tx Signal − Stream 1
0.2 Bits
−0.2
0 5 10 15 20 25 30 35
Time (µ s)
1.2
0.8
Amplitude
0.6
0.4
Tx Signal − Stream 2
0.2 Bits
−0.2
0 5 10 15 20 25 30 35
Time (µ s)
Figure E.2: Beginning of transmit signal for NT 2. Time t 0 µs marks the first bit.
0.8
0.6
0.5
Vout [V]
0.3
0.2
0.1
0
0 2 4 6 8 10 12 14 16 18
Time [ms]
Figure E.3: Oscilloscope measurement for: (a) an unmodulated carrier, (b) no input—noise
only.
162
E.3 sMILE Test Bed Measurements
0.7
Vout
|Vin|
0.6
0.5
0.4
[V]
out
0.3
V
0.2
0.1
0
4950 4960 4970 4980 4990 5000 5010 5020 5030 5040 5050
Time [µ s]
Figure E.4: Oscilloscope output Vpout rns (SISO case) and the computed baseband input volt-
age |Vpin rns|. The latter is not in scale with the former.
1.8721
1.4977
1
nf
|Vin | Rx Signal [mV] 0.7488
0.5
Noise Floor
Rx Bits
|Tx Signal| 0.3744
0 0
0 5 10 15 20 25
Time (µ s)
1.9797
1.5838
1
|Vin| − Rx Signal [mV]
1.1878
Tx Signal
f
|Vin| Rx Signal [mV]
0.7919
0.5 Noise Floor
Rx Bits
|Tx Signal| 0.3959
0 0
0 5 10 15 20 25
Time (µ s)
163
E sMILE Test Bed
140
140
100
120
120
80 100
100
Histogram
Histogram
Histogram
80
Zeros received at −57 dBm 80
60 Zeros received at −67 dBm Zeros received at −76 dBm
Fitted Gaussian distribution
Fitted lognormal distribution Fitted lognormal distribution
Fitted Rician distribution
60 60
40
40 40
20
20 20
0 0 0
0.385 0.39 0.395 0.4 0.405 0.41 0.115 0.12 0.125 0.13 0.135 0.03 0.035 0.04 0.045 0.05 0.055
|V | [mV] |V | [mV] |Vin| [mV]
in in
Figure E.7: Histogram of y1 rms and fitted lognormal distributions for a received signal at
(a) 57 dBm, (b) 67 dBm and (b) 76 dBm.
−3 −3
x 10 x 10
6.5 6.5
sMILE0 σ1(|h|) = 0.004483 |h| + 0.0026
sMILE2
6 sMILE3 6
sMILE4
5.5 5.5
Standard deviation σ1
Standard deviation σ1
5 5
4.5 4.5
4 4
3.5 3.5
3 3
samples
linear fitting
2.5 2.5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
Channel Strength |h| [mV] Channel Strength |h| [mV]
(a) (b)
Figure E.8: (a) Standard deviation of Gaussian approximation for all sMILE nodes as a
function of the channel strength, (b) linear fitting of the standard deviation.
164
E.3 sMILE Test Bed Measurements
300
Zeros received at −57 dBm Zeros received at −67 dBm Zeros received at −76 dBm
Fitted lognormal distribution Fitted lognormal distribution Fitted lognormal distribution
Fitted Rayleigh distribution 250
200 250
200
200
150
Histogram
Histogram
Histogram
150
150
100
100
100
50
50 50
0 0 0
2 4 6 8 10 12 14 16 2 4 6 8 10 12 14 2 4 6 8 10 12 14 16
|V | [mV] −3
x 10 |V | [mV] −3
x 10 |V | [mV] −3
x 10
in in in
Figure E.9: Histogram of y0 rms and fitted lognormal distributions for a received signal at
(a) 57 dBm, (b) 67 dBm and (b) 76 dBm.
1 1
0 0
−1 −1
sMILE0: µ
Parameter value
Parameter Value
sMILE2: µ samples of µ0
−2 sMILE3: µ −2
linear fitting for µ0
sMILE4: µ
sMILE0: σ samples of σ0
−3 sMILE2: σ −3 average of σ
0
sMILE3: σ
sMILE4: σ
−4 −4
−6 −6
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
Channel Strength |h| [mV] Channel Strength |h| [mV]
(a) (b)
Figure E.10: (a) Lognormal parameters pµ0 , σ0 q for all sMILE nodes as a function of the
channel strength, (b) lognormal parameters and fitted curves: average for σ0
and linear fitting for µ0 .
165
E sMILE Test Bed
0.875
0.75
0.65
|h11|
Channel strength [mV]
|h21|
0.5
|h +h | |h21+h22|
0.375 11 12
0.25 |h12| |h |
22
0.125
0
10 20 30 40 50 60 70 80 90 100
Packets
Figure E.11: Sufficient CSI for a 2 2 MIMO system, across all transmitted packets. Trans-
mission is conducted via cables and splitter/combiners and the Racoon’s use
independent LO’s. Npilot 10, NSpP 250.
0.45
|h11+h12|
0.5
0.4
0.35 |h +h |
11 12
0.3 0.375
|h21+h22|
|h | |h |
Channel strength [mV]
11 12
0.25
0.2 0.25
|h11|
|h21| |h22| |h21| |h22|
0.15 |h12|
0.1 0.125
|h21+h22|
0.05
0 0
10 20 30 40 50 60 70 80 90 100 10 20 30 40 50 60 70 80 90 100
Packets Packets
(a) (b)
Figure E.12: Sufficient CSI of a 2 2 MIMO system, across all transmitted packets for two
wireless transmissions (without changing the position of the antennas). The
RACooN’s use independent LO’s. Npilot 10, NSpP 250.
166
E.3 sMILE Test Bed Measurements
0.4
0.35
|h11+h12| |h21+h22|
0.35
0.3
Channel Strength [mV]
0.3
0.2
|h11+h12| |h +h |
0.2 21 22
|h22|
|h11| |h21| |h22|
0.15 |h11| |h12|
0.15
0.1
0.1
0.05 0.05
|h12|
0 0
10 20 30 40 50 60 70 80 90 100 10 20 30 40 50 60 70 80 90 100
Packets Packets
(a) (b)
Figure E.13: Sufficient CSI for a 2 2 MIMO system. RACooN’s share the same LO.
Npilot 10, NSpP 250. (a) Wireless and (b) wired transmission.
0.5
0.35
0.45
0.3 0.4
Channel Strength [mV]
0.35
0.25
0.3
0.2 0.25
0.2
0.15
0.15
0.1
0.1
0.05
0.05 0
10 20 30 40 50 60 70 80 90 100 10 20 30 40 50 60 70 80 90 100
Packets Packets
(a) (b)
0.625
0.4
0.5
Channel Strength [mV]
0.35
Channel Strength [mV]
0.375
0.3
0.25
0.25
0.2
0.125
0.15
0 0.1
10 20 30 40 50 60 70 80 90 100 10 20 30 40 50 60 70 80 90 100
Packets Packets
(c) (d)
Figure E.14: Sufficient CSI of a 4 4 MIMO system. RACooN’s #(1, 2) and #(3, 4) use
the same LO. Npilot 10, NSpP 250. (a) Direct channels, sums of (b) two,
(c) three and (d) four coefficients.
167
E sMILE Test Bed
Tx Signal − Stream 1
0.5 Bits
0
0 10 20 30 40 50 60 70
Time (µ s)
Tx Signal − Stream #2, RACooN #5
1
Amplitude
0
0 10 20 30 40 50 60 70
Time (µ s)
Tx Signal − Stream #3, RACooN #8
1
Amplitude
Tx Signal − Stream 3
0.5 Bits
0
0 10 20 30 40 50 60 70
Time (µ s)
Tx Signal − Stream #4, RACooN #3
1
Amplitude
Tx Signal − Stream 4
0.5
Bits
0
0 10 20 30 40 50 60 70
Time (µ s)
Figure E.15: Transmit signal for a 4x4 MIMO system: synchronization with Barker code,
SEES pilot with Npilot 4.
168
E.3 sMILE Test Bed Measurements
0.8 0.3735
0.6 0.249
0.4
0.1245
0.2
0 0
0 10 20 30 40 50 60 70
Time (µ s)
Input level = −68.2 dBm (Rx #2)
0.3402
0.2041
0.5 0.1361
0.068
0 0
0 10 20 30 40 50 60 70
Time (µ s)
Input level = −60.5 dBm (Rx #3)
0.8275
0.8 0.4965
0.6
0.331
0.4
0.2 0.1655
0 0
0 10 20 30 40 50 60 70
Time (µ s)
Input level = −59.3 dBm (Rx #4)
0.954
|Vin| − Rx Signal [mV]
0.7632
1
Tx Signal
0.8 0.5724
0.6 0.3816
0.4
0.1908
0.2
0 0
0 10 20 30 40 50 60 70
Time (µ s)
Figure E.16: Received signal for a 4 4 MIMO system: synchronization with Barker code,
SEES pilot with Npilot 4.
169
E sMILE Test Bed
|V | − Rx Signal [mV]
0.4981
1
Tx Signal
0.8 0.3735
0.6 0.249
0.4
0.1245
0.2
0 0
80 82 84 86 88 90 92 94 96 98 100
Time (µ s)
Input level = −68.2 dBm (Rx #2)
0.3402
|V | − Rx Signal [mV]
1 0.2721
Tx Signal
0.2041
0.5 0.1361
0.068
0 0
80 82 84 86 88 90 92 94 96 98 100
Time (µ s)
Input level = −60.5 dBm (Rx #3)
0.8275
|V | − Rx Signal [mV]
1 0.662
Tx Signal
0.8 0.4965
0.6
0.331
0.4
0.2 0.1655
0 0
80 82 84 86 88 90 92 94 96 98 100
Time (µ s)
Input level = −59.3 dBm (Rx #4)
0.954
|V | − Rx Signal [mV]
0.7632
1
Tx Signal
0.8 0.5724
0.6 0.3816
0.4
0.2 0.1908
0 0
80 82 84 86 88 90 92 94 96 98 100
Time (µ s)
Figure E.17: Received signal for a 4 4 MIMO system from Fig. E.16.
170
E.3 sMILE Test Bed Measurements
|V | − Rx Signal [mV]
0.0465
1
Tx Signal
0.8 0.0349
0.6 0.0233
0.4
0.0116
0.2
0 0
0 10 20 30 40 50 60 70 80 90 100
Time (µ s)
Input level = −62.2 dBm (Rx #2)
0.6454
|V | − Rx Signal [mV]
0.5163
1
Tx Signal
0.8 0.3873
0.6 0.2582
0.4
0.1291
0.2
0 0
0 10 20 30 40 50 60 70 80 90 100
Time (µ s)
Figure E.18: Received signal for a 2 2 MIMO system. The received signal on the first
antenna is 21 dB weaker than on the second antenna.
Figure E.19: Two clusters of two RACooN nodes each are depicted in the corridor of ETF
F floor. These positions were used for the the mixed LOS-NLOS measurement
of Table 6.4 and Fig. E.20b.
171
E sMILE Test Bed
Tx
xx xx
Corridor
Corridor Tx1
Tx2 xx xx
4.5 m
ETF F109 ETF F109
4.3 m
xxxx
Rx
xxxx
xx
Rx
Figure E.20: Blueprint of ETF F floor for the measurements in Table 6.4. (a) LOS measure-
ments: the RACooN’s are positioned at distances d 10, 20, 25, 37 m along
the corridor. (b) NLOS and mixed measurements: the two NLOS measure-
ments are conducted with Tx1 (RACooN’s 4 and 5) and Tx2 (RACooN’s 3
and 8) next to each other. For the mixed measurement, Tx2 is moved to the
second position and faces the receiver through the open door of F110.
172
Notation
173
Notation
|S | Cardinality of set.
S Modulation alphabet.
C Set of complex numbers.
N Set of natural numbers.
R Set of real numbers.
174
Acronyms
CF Crest Factor.
CPM Constant Phase Modulation.
CRLB Cramér-Rao Lower Bound.
CSI Channel State Information.
175
Acronyms
I/O Input/Output.
I/Q In-Phase/Quadrature Phase.
IC Integrated Circuit.
IF Intermediate Frequency.
ISI Intersymbol Interference.
176
Acronyms
177
Acronyms
178
Bibliography
179
Bibliography
180
Bibliography
[23] ——, “sMILE: The first MIMO Amplitude Detection Test Bed,” in to be submitted,
2011.
[24] ——, “Channel Estimation for Very Low Power MIMO Envelope Detectors,” IEEE
Transactions on Wireless Communications, vol. 9, pp. 824–832, Feb. 2010.
[25] ——, “Nonlinear MIMO: Affordable MIMO Technology for Wireless Sensor Net-
works,” to be submitted in a Journal.
[26] G. Fettweis, M. Löhning, D. Petrovic, M. Windisch, P. Zillmann, and W. Rave,
“Dirty RF: A New Paradigm,” International Journal of Wireless Information Net-
works, vol. 14, pp. 133–148, June 2007.
[27] A. Papoulis, Probability, Random Variables, and Stochastic Processes, 3rd ed.
McGraw-Hill, 1991.
[28] W. R. Bennett, “Methods of solving noise problems,” in Proceedings of the IRE,
vol. 44, May 1956, pp. 609–638.
[29] R. Pawula, S. Rice, and J. Roberts, “Distribution of the Phase Between Two
Vectors Perturbed by Gaussian Noise,” IEEE Transactions on Communications,
vol. 30, pp. 1828–1841, Aug. 1982.
[30] S. Weinzierl, “Introduction to Monte Carlo Methods,” in Arxiv, Rep. Nr. NIKHEF-
00-012, June 2000.
[31] K. S. Miller, Multidimensional Gaussian Distributions. Wiley, 1964.
[32] R. K. Mallik, “On Multivariate Rayleigh and Exponential Distributions,” in IEEE
Transactions on Information Theory, vol. 49, no. 6, June 2003, pp. 1499–1515.
[33] Y. Chen and C. Tellambura, “Infinite Series Representaion of the Trivariate and
Quadrivariate Rayleigh Distribution and Their Applications,” IEEE Transactions
on Communications, vol. 53, pp. 2092–2101, Dec. 2005.
[34] M. Abramowitz and I. Stegun, Handbook of mathematical functions with formulas,
graphs, and mathematical table, 1965.
[35] T. Yo and A. Goldsmith, “Capacity and Power Allocation for Fading MIMO Chan-
nels With Channel Estimation Error,” in IEEE Transactions on Information The-
ory, vol. 52, no. 5, May 2006, pp. 2203–2214.
[36] J. P. Aldis and A. G. Burr, “The Channel Capacity of Discrete Time Phase Modu-
lation in AWGN,” IEEE Transactions on Information Theory, vol. 39, pp. 184–185,
Jan. 1993.
181
Bibliography
[37] K. Azarian and H. El Gamal, “On the Utility of a 3 db SNR Gain in MIMO
Channels,” in IEEE International Symposium on Information Theory, July 2006,
pp. 2274–2278.
[38] E. H. Calhoun, D. C. Daly, N. Verma, D. F. Finchelstein, D. D. Wentzloff,
A. Wang, S.-H. Cho, and A. P. Chandrakasan, “Design Considerations for Ultra-
Low Energy Wireless Microsensor Nodes,” IEEE Transactions on Computers,
vol. 54, pp. 727–740, June 2005.
[39] J. G. Proakis, Digital Communications, 4th ed. McGraw-Hill, 2001.
[40] M. Schwartz, W. R. Bennett, and S. Stein, Communication Systems and Tech-
niques. IEEE Press, 1995.
[41] J. Geist, “Asymptotic error rate behavior for noncoherent on-off keying,” IEEE
Transactions on Communications, vol. 42, p. 225, Jan 1994.
[42] E. Biglieri, Coding for Wireless Channels. Springer, 2005.
[43] C. Tepedelenlioğlu, A. Abdi, and G. Giannakis, “The Ricean K Factor: Estima-
tion and Performance Analysis,” IEEE Transactions on Wireless Communications,
vol. 2, no. 4, pp. 799–810, July 2003.
[44] K. K. Talukdar and W. D. Lawing, “Estimation of the parameters of the Rice
distribution,” The Journal of the Acoustical Society of America, pp. 1193–1197,
Mar. 1991.
[45] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation Theory.
Englewood Cliffs, NJ: Prentice Hall, 1993.
[46] M. Biguesh and A. Gershman, “Training-Based MIMO Channel Estimation: A
Study of Estimator Tradeoffs and Optimal Training Signals,” IEEE Transactions
on Signal Processing, vol. 54, no. 3, pp. 884–893, Mar. 2006.
[47] S. Haykin, Communication Systems, 2nd ed. Wiley, 1983.
[48] A. Devices, Datasheet: Fast, Voltage-Out DC-440 MHz, 95 dB Logarith-
mic Amplifier AD8310. http://www.analog.com/static/imported-files/data_
sheets/ad8310.pdf, 2005.
[49] A. V. Oppenheim and R. W. Schafer, Discrete-Time Signal Processing, 2nd ed.
Prentice-Hall, 1999.
[50] S. Berger, Dissertation: Coherent Cooperative Relaying in Low Mobility Wireless
Multiuser Networks. Logos Verlag Berlin, 2010.
182
Bibliography
[51] E. Veach, Dissertation: Robust Monte Carlo Methods for Light Transport Simula-
tion. University of Stanford Dissertations, 1997.
[52] T. M. Cover and J. A. Thomas, Elements of information theory. New York, NY,
USA: Wiley-Interscience, 1991.
183
Curriculum Vitae
Education
11/2005–01/2011 Swiss Federal Institute of Technology (ETH) Zurich,
Switzerland, PhD studies at the Department of Information
Technology and Electrical Engineering, Communication Tech-
nology Laboratory.
10/2003–10/2005 Technische Universität München, Germany, MSc. at
the Electrical Engineering and Information Technology De-
partment, Communications Engineering.
09/1998–07/2003 Aristotle University of Thessaloniki, Greece, Diploma
at the Electrical Engineering and Computer Engineering
Department.
Scholarship
11/2004 Stibet scholarship, sponsored by the German Academic Ex-
change Service in recognition of outstanding results in the
MSCE program (MSc.).
Experience
11/2005–01/2011 ETH Zurich, Switzerland, Research Assistant at Commu-
nication Technology Laboratory, Prof. Dr. A. Wittneben.
185
Patents
1. C. Esli, G. K. Psaltopoulos, J. Wagner and A. Wittneben, User cooperation
aided multitasking for cellular communications. Disclosure with NOKIA, Radio
Implementation Patent Board, Finland, 2008.
2. J. Wagner, G. K. Psaltopoulos, C. Esli and A. Wittneben, Intra-virtual an-
tenna array communication through secondary MIMO systems beyond 10 GHz.
Disclosure with NOKIA, Radio Implementation Patent Board, Finland, 2008.
Publications
1. G. K. Psaltopoulos and A. Wittneben, "sMILE: the first MIMO Evelope De-
tection test bed", IEEE Vehicular Technology Conference, submitted 2011.
2. G. K. Psaltopoulos and A. Wittneben, "Channel Estimation for Very Low
Power MIMO Envelope Detectors", IEEE Transactions on Vehicular Technology,
submitted 2011.
3. G. K. Psaltopoulos and A. Wittneben, "Nonlinear MIMO: Affordable MIMO
Technology for Wireless Sensor Networks", IEEE Transactions on Wireless Com-
munications, Vol. 9, Iss. 2, pp. 824-832, Feb. 2010.
4. G. K. Psaltopoulos and A. Wittneben, "Channel Estimation for Very Low
Power MIMO Envelope Detectors", IEEE International Conference on Commu-
nications, ICC 2010, Cape Town, May 2010
5. G. K. Psaltopoulos and A. Wittneben, "Diversity and Spatial Multiplexing of
MIMO Amplitude Detection Receivers", IEEE International Symposium on Per-
sonal, Indoor and Mobile Radio Communications, PIMRC 2009, Tokyo, Septem-
ber 2009
6. G. K. Psaltopoulos and A. Wittneben, "Achievable Rates of Nonlinear MIMO
Systems with Noisy Channel State Information", IEEE International Symposium
on Information Theory, ISIT 2008, Toronto, July 2008
7. G. K. Psaltopoulos, F. Trösch and A. Wittneben, "On Achievable Rates of
MIMO Systems with Nonlinear Receivers", IEEE International Symposium on
Information Theory, ISIT 2007, Nice, June 2007
8. G. K. Psaltopoulos, M. Joham and W. Utschick, "Generalized MMSE Detec-
tion Techniques for Multipoint-to-Point Systems", IEEE Global Telecommunica-
tions Conference, Globecom 2006, San Francisco, November 2006
9. G. K. Psaltopoulos, M. Joham and W. Utschick, "Comparison of Lattice
Search Techniques for Nonlinear Precoding", International ITG/WSA Workshop
on Smart Antennas, Ulm, Germany, March 2006
186